1<!DOCTYPE html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width, initial-scale=1"><title>What can 500 years of journalism teach developers about AI trustworthiness? - Stack Overflow</title><link rel="preconnect" href="https://www.googletagmanager.com">
1<script type="importmap">{"imports":{"#entry":"/_nuxt/igBnPZTB.js"}}</script>
vendor: 57 bytes, line 1
1<script src="https://www.googletagmanager.com/gtag/js?id=
1G-DTTX7JCXEJ
vendor: 36 bytes, line 1
1" async data-hid="head-ga"></script>
1<style>.p-article table{border-collapse:collapse;border-spacing:0;margin:2em auto;overflow:scroll;word-break:normal;word-break:keep-all;-webkit-overflow-scrolling:touch;width:100%}.p-article table th{font-weight:700}.p-article table td,.p-article table th{border:1px solid var(--black-100);padding:6px 13px}.p-article table tr:nth-child(2n){background-color:var(--black-100)}.p-article table tr:hover{background-color:var(--blue-200)}.p-article table ol,.p-article table ul{margin:0;padding:0}.p-article .wp-block-image{margin:2em 0;text-align:center}.p-article figcaption{color:var(--black-500);font-size:.8em;line-height:1.3em;margin:25px auto 0;max-width:620px}</style><style>nav a[data-v-df5a0bb6],nav a[data-v-df5a0bb6]:visited,nav button[data-v-df5a0bb6]{color:var(--black)!important;display:block;padding:3px 0}nav a[data-v-df5a0bb6]:hover,nav button[data-v-df5a0bb6]:hover{color:var(--blue-500);text-decoration:underline}nav li>ul[data-v-df5a0bb6]{border-left:3px solid var(--black-225);margin-top:3px;padding-left:10px}.sub-link[data-v-df5a0bb6]{color:var(--black-400)!important}section h3[data-v-df5a0bb6]{color:var(--black-400);font-size:16px;font-weight:500;margin:12px 0 24px}@media screen and (max-width:470px){.sections[data-v-df5a0bb6]{flex-direction:column}}</style><style>.header[data-v-7442e526]{position:relative}.header a[data-v-7442e526]{opacity:0;padding:0 12px;position:absolute;right:100%}.header a[data-v-7442e526]:before{content:"#"}.header:hover a[data-v-7442e526]{opacity:1}</style><link rel="stylesheet" href="/_nuxt/entry.DB50Bfqr.css" crossorigin><link rel="stylesheet" href="/_nuxt/Footer.CBW8JzNG.css" crossorigin><link rel="stylesheet" href="/_nuxt/Modal.CnPlmjv2.css" crossorigin><link rel="stylesheet" href="/_nuxt/SanityContent.WzRGkE7M.css" crossorigin><link rel="preload" as="fetch" crossorigin="anonymous" href="/2026/06/08/what-can-500-years-of-journalism-teach-developers-about-ai-trustworthiness/_payload.json?_b=4900ffdd-7672-4b60-9bc4-4d8cbbb19c70"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/igBnPZTB.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/hePW80VL.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/DjS0xbvg.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/CNW9Osv5.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/HclGiUj8.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/D1qpH6Pu.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/ULRhyFxa.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/BGze3w87.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/Ccl0v3aC.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/BaUt5Xm8.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/6lvfEp2t.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/CkvdZEJQ.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/mHzno_Pf.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/C3gn72eF.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/63w8-Y63.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/DfowE93g.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/CRfAUoyq.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/C_xuQalc.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/CVKolL_72.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/DUTxQXBQ.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/DJrHMbtv.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/C1msyEQU.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/D2GSWYk82.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/DjQPAAOC2.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/D-ERr1vr.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/BTfY3gI02.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/70_x2W4k2.js"><link rel="modulepreload" as="script" crossorigin href="/_nuxt/B8TO32Ue2.js">
1<script src="https://cdn.cookielaw.org/scripttemplates/otSDKStub.js" data-domain-script="d06435e0-17fb-4659-9a14-1930a6e0be80-test" defer data-hid="head-onetrust-sdk"></script>
1<script type="module" src="/_nuxt/igBnPZTB.js" crossorigin></script>
1<link rel="dns-prefetch" href="https://www.google-analytics.com"><link rel="prefetch" as="script" crossorigin href="/_nuxt/zxfBb7ct.js"><link rel="prefetch" as="script" crossorigin href="/_nuxt/DFd1IrEG.js"><link rel="prefetch" as="script" crossorigin href="/_nuxt/Cx3v2-Yp.js"><link rel="prefetch" as="script" crossorigin href="/_nuxt/ODC8zPNr.js"><link rel="prefetch" as="script" crossorigin href="/_nuxt/DQT0gUW1.js"><link rel="prefetch" as="script" crossorigin href="/_nuxt/CuX0iEMe.js"><link rel="prefetch" as="script" crossorigin href="/_nuxt/Datf-Mgm.js"><link rel="prefetch" as="script" crossorigin href="/_nuxt/B9kHqgaw.js"><link rel="prefetch" as="script" crossorigin href="/_nuxt/wzadwL2c.js"><link rel="prefetch" as="script" crossorigin href="/_nuxt/BXR6Gh-7.js"><meta name="format-detection" content="telephone=no"><meta name="theme-color" content="#FF5E00"><meta name="twitter:site" content="@stackoverflow"><meta name="twitter:card" content="summary_large_image"><meta name="robots" content="Index, Follow"><meta name="description" content="AI reliability issues stem from three separate architectural challenges that keep getting lumped into the same category. Prompt engineering alone can't fix them. But the sourcing and verification frameworks media organizations have used for centuries translate into clear engineering solutions developers can implement today."><meta property="og:type" content="website"><link rel="icon" href="/favicon.ico" sizes="any"><link rel="icon" href="/favicon.svg" type="image/svg+xml"><link rel="apple-touch-icon" href="/apple-touch-icon.png"><link rel="image_src" href="/apple-touch-icon.png"><meta property="og:url" content="https://stackoverflow.blog/2026/06/08/what-can-500-years-of-journalism-teach-developers-about-ai-trustworthiness/"><link rel="canonical" href="https://stackoverflow.blog/2026/06/08/what-can-500-years-of-journalism-teach-developers-about-ai-trustworthiness/"><meta property="og:image" content="https://cdn.stackoverflow.co/images/jo7n4k8s/production/add2055954555b916ecd60384611e44e891d64af-12000x6300.jpg?w=1200&h=630&auto=format"></head><body class="">
1<script data-hid="body-theme-select">let stackex_theme = localStorage.getItem('stackex-theme'); if (stackex_theme) document.body.className += ' ' + stackex_theme</script>
1<div id="__nuxt"><!--[--><!--[--><div class="nuxt-loading-indicator" style="position:fixed;top:0;right:0;left:0;pointer-events:none;width:auto;height:3px;opacity:0;background:#FF5E00;background-size:0% auto;transform:scaleX(0%);transform-origin:left;transition:transform 0.1s, height 0.4s, opacity 0.4s;z-index:999999;"></div><!--[--><!--[--><header class="bb bc-black-300 py12 ps-relative z-nav"><div class="p-wrapper d-flex w100 mx-auto ff-sans"><div class="flex--item4"><a href="/" class="d-inline-flex ai-center h:bg-black-150 bar-md p6"><svg width="32" height="32" viewBox="0 0 32 32" class="svg-icon IconGlyph32 fc-black-600" aria-hidden="true"><path d="M23.8 17.22a18 18 0 0 0-2.2 3.87l-.02.06q-.83 2.04-1.15 4.3v.03A18 18 0 0 0 20.4 30H4v-4.55h14.23l.1-.56L4.6 21.1l1.14-4.4 13.88 3.85.2-.44-12.46-7.43 2.2-3.94 12.65 7.54.3-.37L12.17 5.2 15.28 2l10.53 10.86 1.1 1.15q-1.75 1.4-3.11 3.21"></path></svg><span class="fc-black-500 fw-bold fs-body3 bl ml12 pl12 bc-black-200">Blog</span></a></div><form class="print:d-none flex--item as-center ml-auto mr12 md:mr0"><div><div class="ps-relative sm:ps-absolute t100 l0 r0 mt1 sm:d-none"><div id="search" class="bg-black-100 sm:pl12 sm:pt12 sm:pr12 sm:pb12"><div class="ps-relative"><input placeholder="Find somethingâ¦" class="s-input s-input__search fs-body2" type="text" autocomplete="off" spellcheck="false"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconSearch s-input-icon s-input-icon__search" aria-hidden="true"><path d="M8.25 1a7.25 7.25 0 0 1 6.67 10.1l-1.17-1.17a5.75 5.75 0 1 0-.94 1.82l1.07 1.07 4.15 4.15-1.06 1.06-4.15-4.15A7.25 7.25 0 1 1 8.25 1"></path></svg><!----></div></div><!----></div><div class="d-none sm:d-block"><button type="button" class="s-btn__clear s-topbar--item p12 s-btn s-btn__icon" role="menuitem" aria-label="Search" aria-haspopup="true" aria-controls="search" title="Click to show search"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconSearch" aria-hidden="true"><path d="M8.25 1a7.25 7.25 0 0 1 6.67 10.1l-1.17-1.17a5.75 5.75 0 1 0-.94 1.82l1.07 1.07 4.15 4.15-1.06 1.06-4.15-4.15A7.25 7.25 0 1 1 8.25 1"></path></svg></button></div></div></form><div class="flex--item d-flex jc-end"><div class="sm:d-none d-flex jc-end fl1 mr24"><a class="s-btn s-btn__clear p8" href="https://stackoverflow.com"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconGlyph" aria-hidden="true"><path d="M14.83 10.7a10 10 0 0 0-1.31 2.21l-.02.03q-.5 1.17-.68 2.46v.02a10 10 0 0 0-.03 2.58H3v-2.6h8.5l.06-.32-8.2-2.17.69-2.5 8.29 2.19.12-.25L5 8.1l1.32-2.26 7.55 4.3.19-.2-6.2-6.11L9.75 2l6.29 6.2.66.66a10 10 0 0 0-1.86 1.84"></path></svg></a><a class="s-btn s-btn__clear p8" href="/feed"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconRSS" aria-hidden="true"><path d="M4 14a2 2 0 1 1 0 4 2 2 0 0 1 0-4M2.53 7.53A10.5 10.5 0 0 1 12.48 18h-1.5A9 9 0 0 0 2 9.02v-1.5zM2.4 2.02A16 16 0 0 1 18 18h-1.5c0-8-6.5-14.48-14.49-14.49v-1.5z"></path></svg></a><a class="s-btn s-btn__clear p8" href="https://stackoverflow.com/users/email/settings/current"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconMail" aria-hidden="true"><path d="M16.5 4c.78 0 1.42.59 1.5 1.35v9.15h-1.5V16H3.35A1.5 1.5 0 0 1 2 14.64L2 14.5V5.35A1.5 1.5 0 0 1 3.36 4L3.5 4zm-5.04 6.73c-.84.72-2.08.72-2.92 0L3.5 6.42v8.08h13V6.42zM9.51 9.6c.28.24.7.24.98 0l4.77-4.09H4.74z"></path></svg></a></div><div class="flex--item fl-shrink0 s-spinner s-spinner__sm as-center fc-black-300"><div class="v-visible-sr">Loadingâ¦</div></div></div></div></header><!----><!--]--><div class="p-wrapper bbr-lg mx-auto bg-white pt32 pb64 sm:pt12"><div class="d-flex lg:fd-column"><div class="flex--item2 ps-relative lg:d-none mr48 sm:mr0"><div><nav class="fs-subheading"><ul class="list-reset sm:d-flex sm:fw-wrap"><li class="my4"><a href="/" class="px12 py6 d-block bar-pill d-flex ai-center ws-nowrap fc-black-500 h:bg-black-200"><svg width="32" height="32" viewBox="0 0 32 32" class="svg-icon IconBlog32 mr12" aria-hidden="true"><path d="M28 2a2.5 2.5 0 0 1 2.5 2.5v23l-.01.26A2.5 2.5 0 0 1 28 30H7a5 5 0 0 1-5-4.74V16h7.52V4.5a2.5 2.5 0 0 1 2.5-2.5zM4.5 25a2.5 2.5 0 0 0 5 0v-6.5h-5zm7.52 2.5H28v-23H12.02zM26 25.06H14v-2h12zm0-4H14v-2h12zM26 17H14v-2h12zm-6-4h-6V7h6z"></path></svg> Everything </a></li><li class="my4"><a href="/contributed/" class="px12 py6 d-block bar-pill d-flex ai-center ws-nowrap fc-black-500 h:bg-black-200"><svg width="32" height="32" viewBox="0 0 32 32" class="svg-icon IconUser32Stack mr12" aria-hidden="true"><path d="M15.78 17c.9.07 1.69.64 2.04 1.48l3 7.2a2.4 2.4 0 0 1-2.06 3.31l-.16.01H5.8a2.4 2.4 0 0 1-2.4-2.5h15.05l-2.91-7H8.87L6.99 24H4.3l2.3-5.52A2.4 2.4 0 0 1 8.8 17h6.98m6.02 0c1.52 0 2.89.95 3.42 2.37l1.75 4.7A3.65 3.65 0 0 1 23.55 29h-.24v-2.5h.24c.8 0 1.35-.8 1.08-1.55l-1.75-4.7c-.17-.45-.6-.75-1.08-.75H20V17zM12.52 3.2a6 6 0 0 1 5.67 5.97v.31a6 6 0 0 1-5.97 5.67l-.31-.01a6 6 0 0 1-5.66-5.66v-.3c0-3.3 2.67-5.98 5.97-5.98zM20 3.19a5.99 5.99 0 0 1 0 11.97v-2.5a3.49 3.49 0 0 0 0-6.97zM12.22 5.7a3.48 3.48 0 0 0 0 6.95 3.48 3.48 0 0 0 0-6.95"></path></svg> The Heap </a></li><li class="my4"><a href="/productivity/" class="px12 py6 d-block bar-pill d-flex ai-center ws-nowrap fc-black-500 h:bg-black-200"><svg class="svg-icon mr12" height="32" viewBox="0 0 32 32" width="32" aira-hidden="true"><path d="m15.1689 17.7533-1.7685 1.7676-2.7998-2.7998-7.49318 7.4941-1.76758-1.7675 8.37696-8.378.8848-.8838zm14.8301 1.2471h-2.5v-6.2891l-10.4541 10.4541-1.7685-1.7685 1.7685-1.7676 9.1299-9.1289h-7.1748v-2.5h10.999z" fill="currentColor"></path></svg> Productivity </a></li><li class="my4"><a href="/ai/" class="px12 py6 d-block bar-pill d-flex ai-center ws-nowrap fc-black-500 h:bg-black-200"><svg width="32" height="32" viewBox="0 0 32 32" class="svg-icon IconAssistant32 mr12" aria-hidden="true"><path d="M13 4.5H7.75c-.69 0-1.25.56-1.25 1.25v5.6a3.75 3.75 0 0 1-3 3.67v2.04c1.71.35 3 1.86 3 3.68v5.51c0 .69.56 1.25 1.25 1.25H13V30H7.75A3.75 3.75 0 0 1 4 26.25v-5.51c0-.7-.56-1.25-1.25-1.25H1v-6.9h1.75c.69 0 1.25-.56 1.25-1.25V5.75A3.75 3.75 0 0 1 7.75 2H13zM24.25 2A3.75 3.75 0 0 1 28 5.75v5.59c0 .69.56 1.24 1.25 1.25H31v6.9h-1.75c-.69 0-1.25.55-1.25 1.24v5.52A3.75 3.75 0 0 1 24.25 30H19v-2.5h5.25c.7 0 1.25-.56 1.25-1.25v-5.52a3.75 3.75 0 0 1 3-3.67V15a3.75 3.75 0 0 1-3-3.67V5.75c0-.69-.56-1.25-1.25-1.25H19V2zM13 16h-3v-3h3zm9 0h-3v-3h3z"></path></svg> AI/ML </a></li><li class="my4"><a href="/open-source/" class="px12 py6 d-block bar-pill d-flex ai-center ws-nowrap fc-black-500 h:bg-black-200"><svg class="svg-icon mr12" height="32" viewBox="0 0 32 32" width="32" aira-hidden="true"><path d="m16.4121 1.48145c3.7482.19029 6.8135 2.96082 7.46 6.57031l-2.4082.79785c-.3088-2.74485-2.6366-4.87881-5.4639-4.87891-3.0373.0001-5.4997 2.46272-5.5 5.5v4h8.7002v.0293h9.7998v16.5h-26v-16.5h5v-4.02832l.01074-.41309c.21473-4.22646 3.70946-7.58877 7.98926-7.58886zm-10.9121 26.01855h21v-11.5h-21z" fill="currentColor"></path></svg> Open Source </a></li><li class="my4"><a href="/business/" class="px12 py6 d-block bar-pill d-flex ai-center ws-nowrap fc-black-500 h:bg-black-200"><svg class="svg-icon mr12" height="32" viewBox="0 0 32 32" width="32" aira-hidden="true"><path d="m21.4998 4.49561h2.5v7.50489h4.9942v18.0117h-4.9961v-.0117l-15.99808-.001h-5v-16h5v-11.99901h13.49998zm2.5 23.01659h2.4942v-13.0117h-2.4942zm-13.5-.0117h11v-23.00001h-11zm-4.99998-.001h2.5v-11h-2.5zm13.49998-5.499h-6v-2h6zm0-4h-6v-2h6zm0-4h-6v-2h6zm0-4h-6v-2.00001h6z" fill="currentColor"></path></svg> Business Hub <!----></a></li></ul><ul class="list-reset pt12 mt16 bt bc-black-200 sm:d-flex sm:fw-wrap"><li class="my4"><a href="/company/" class="px12 py6 d-block bar-pill d-flex ai-center ws-nowrap fc-black-500 h:bg-black-200"><svg width="32" height="32" viewBox="0 0 32 32" class="svg-icon IconGlyph32 mr12" aria-hidden="true"><path d="M23.8 17.22a18 18 0 0 0-2.2 3.87l-.02.06q-.83 2.04-1.15 4.3v.03A18 18 0 0 0 20.4 30H4v-4.55h14.23l.1-.56L4.6 21.1l1.14-4.4 13.88 3.85.2-.44-12.46-7.43 2.2-3.94 12.65 7.54.3-.37L12.17 5.2 15.28 2l10.53 10.86 1.1 1.15q-1.75 1.4-3.11 3.21"></path></svg> Company </a></li><li class="my4"><a href="/releases/" class="px12 py6 d-block bar-pill d-flex ai-center ws-nowrap fc-black-500 h:bg-black-200"><svg width="32" height="32" viewBox="0 0 32 32" class="svg-icon IconStar32 mr12" aria-hidden="true"><path d="M19.6 12.03h11.65l-9.42 6.85.98 3.04-3.44-2.5-.48-1.5 1.47-1.06 3.2-2.33h-5.78l-.56-1.73L16 9.04l-1.22 3.76-.56 1.73H8.44l3.2 2.33 1.47 1.06-.56 1.73-1.22 3.76 3.2-2.32 1.44-1.05 8.16 5.93 1.3 3.99L16 23.1l-9.43 6.85 3.6-11.08-9.42-6.85H12.4L16 .95z"></path></svg> Releases </a></li></ul><ul class="list-reset pt12 mt16 bt bc-black-200 sm:d-flex sm:fw-wrap"><li class="my4"><a href="/podcast/" class="px12 py6 d-block bar-pill d-flex ai-center ws-nowrap fc-black-500 h:bg-black-200"><svg width="32" height="32" viewBox="0 0 32 32" class="svg-icon IconSound32 mr12" aria-hidden="true"><path d="M26.31 5.69a14.6 14.6 0 0 1 0 20.62l-1.77-1.77a12.1 12.1 0 0 0 0-17.08zM16 7v18l-2.14.88L8.04 20H4.5v-2.5H2V12h6.04l5.82-5.88zm-6.56 7.13c-.23.24-.55.37-.88.37H4.5v3h4.18q.45.05.76.37l4.06 4.1V10.03zM21.6 10.4a7.9 7.9 0 0 1 0 11.2l-1.76-1.78a5.4 5.4 0 0 0 0-7.66z"></path></svg> Podcast </a></li><li class="my4"><a href="/newsletter/" class="px12 py6 d-block bar-pill d-flex ai-center ws-nowrap fc-black-500 h:bg-black-200"><svg width="32" height="32" viewBox="0 0 32 32" class="svg-icon IconMail32Stack mr12" aria-hidden="true"><path d="M2.53 24.39c0 .63.51 1.15 1.15 1.15h20.34v2.5H3.68a3.65 3.65 0 0 1-3.65-3.65V10h2.5zM28.52 4c1.3 0 2.36.98 2.48 2.24l.02.26v13.94h-2.5V7.94l-8.18 6.8a3.65 3.65 0 0 1-4.67 0L7.5 7.94v12.51h21v2.5H7.4l-.25-.01a2.4 2.4 0 0 1-2.14-2.14L5 20.55V6.5l.01-.26A2.5 2.5 0 0 1 7.5 4zm-11.25 8.82c.43.35 1.05.35 1.48 0l7.58-6.32H9.7z"></path></svg> Newsletter </a></li></ul></nav><div class="mt16 bt bc-black-200 pl12 pr12 pt24 md:d-none"><div class="tt-uppercase fc-light fs-body1 mb12"><a class="fc-light h:fc-black-500" href="https://stackoverflow.co/"> Stack Overflow Business </a></div><a class="d-block pb12 fw-light fc-black-500 h:fc-blue-500" href="https://stackoverflow.co/internal/?utm_source=so-owned&utm_medium=blog&utm_campaign=nav-side-bar"><div class="fc-black-500 wmx2"><strong class="fw-medium">Stack Internal</strong>: the knowledge intelligence layer that powers enterprise AI.</div></a><a class="d-block pb12 fw-light fc-black-500 h:fc-blue-500" href="https://stackoverflow.co/data-licensing/?utm_source=so-owned&utm_medium=blog&utm_campaign=nav-side-bar"><div class="fc-black-500 wmx2"><strong class="fw-medium">Stack Data Licensing</strong>: decades of verified, technical knowledge to boost AI performance and trust.</div></a><a class="d-block pb12 fw-light fc-black-500 h:fc-blue-500" href="https://stackoverflow.co/advertising/employer-branding/?utm_source=so-owned&utm_medium=blog&utm_campaign=nav-side-bar"><div class="fc-black-500 wmx2"><strong class="fw-medium">Stack Ads</strong>: engage developers where it matters â in their daily workflow.</div></a></div></div></div><article class="post flex--item7 wmn0 lg:wmx100 lg:pb48" itemscope itemtype="https://schema.org/Article"><header class="mb32 pt12"><time datetime="2026-06-08T14:00:00.000Z" class="flex--item tt-uppercase fw-bold fs-body1 fc-black-500 d-flex mb16" itemprop="datePublished"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconCalendar va-middle mr6 mtn1" aria-hidden="true"><path d="M7.6 2h4.8V.6H14V2h4v16H2V2h4V.6h1.6zM3.5 16.5h13v-8h-13zm4-2H6V13h1.5zm3.25 0h-1.5V13h1.5zm3.25 0h-1.5V13H14zm-6.5-3H6V10h1.5zm3.25 0h-1.5V10h1.5zm3.25 0h-1.5V10H14zM3.5 7h13V3.5H14v1.89h-1.6V3.5H7.6v1.89H6V3.5H3.5z"></path></svg> June 8, 2026</time><h1 class="fs-display2 lh-xs ff-stack-sans-headline mb24" itemprop="name">What can 500 years of journalism teach developers about AI trustworthiness?</h1><p class="fs-title fc-black-500 wmx6" itemprop="abstract">AI reliability issues stem from three separate architectural challenges that keep getting lumped into the same category. Prompt engineering alone can't fix them. But the sourcing and verification frameworks media organizations have used for centuries translate into clear engineering solutions developers can implement today.</p><figure class="m0"><img src="https://cdn.stackoverflow.co/images/jo7n4k8s/production/add2
1055954555b916ecd60384611e44e891d64af-12000x6300.jpg?rect=8,0,11985,6300&w=780&h=410&auto=format&dpr=2" width="1200" height="630" class="bar-md w100 mt32 h-auto d-block as-start ba bc-black-150" alt="" itemprop="image"><figcaption class="fc-black-500 pt12"><svg width="16" height="16" viewBox="0 0 16 16" class="svg-icon IconInfo16 mr4" aria-hidden="true"><path d="M8 1a7 7 0 1 1 0 14A7 7 0 0 1 8 1m0 1.25a5.75 5.75 0 1 0 0 11.5 5.75 5.75 0 0 0 0-11.5M8.02 7c.54 0 .98.44.98.98v3.04a.98.98 0 0 1-1.96 0V7.98c0-.54.44-.98.98-.98M8 4a1 1 0 1 1 0 2 1 1 0 0 1 0-2"></path></svg> <span> Credit: Alexandra Francis</span></figcaption></figure></header><div itemprop="articleBody" class="s-prose fs-body2"><!--[--><p>Every time a user queries an AI search engine for information, they're trusting a system trained on the internet to behave like an editor. An editor has institutional memory, a corrections policy, and journalistic accountability. LLMs have none of those things, which isn't just a problem for newsrooms.</p><p>Developers integrating LLMs into documentation tools, research assistants, knowledge bases, and coding copilots rely on output accuracy. When accuracy fails downstream, the consequences are operational: support tickets, compliance gaps, eroded user trust, and, in high-stakes domains like legal or medical tech, real liability.</p><p>Moreover, AI enthusiasts talk about LLMs "getting things wrong" as if it's one problem. It's actually three:</p><ul><li>Unintentional fabrication (hallucinations of both the overconfident and underconfident variety),</li><li>Sycophancy with user prompts, and</li><li>Intentional deception during model evaluation.</li></ul><p>These are structurally distinct failure modes, each caused by something that requires a different fix. Collapsing them under "hallucination" produces mitigations that solve one problem while leaving the others untouched.</p><p>Luckily, journalists have been running information-critical systems long enough to have made these same mistakes, named them, and built institutional responses to each. Those responses were designed to prevent specific operational failures, not serve as abstract ethical guardrails. In fact, most of them translate directly into engineering solutions that apply to any information system, including LLMs.</p><h2 class="header" id="h2-d5bac643fc6e" _key="0aeddc6531fb" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h2-d5bac643fc6e" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Why "hallucination" isn't just one problem</strong><!--]--></span></h2><p>The word "hallucination" has become a catch-all for whenever an LLM says something wrong. That's like calling every aircraft incident a "crash." Itâs too generalized to be useful when discussing prevention.</p><p>Thankfully, there are now multiple studies that have helped us establish a clear distinction between the different ways that large language models can get things wrong. Depending on the underlying engineering issue that causes it, LLM misinformation can be categorized into one of three buckets:</p><h3 class="header" id="h3-46bba7cde39e" _key="a6f146c4df0f" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-46bba7cde39e" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Epistemological mismatch</strong><!--]--></span></h3><p>This is when LLMs canât architecturally distinguish between retrieved knowledge and training-data plausibility. Because fluency and truth-tracking are treated as independent objectives, everything produces equally confident responses by default.</p><p>As a result, attributed claims get silently converted into universal assertions. "Company X reported profits rose" becomes "Profits rose" because nothing in the modelâs architecture penalizes the omission of details for the sake of brevity. Northwestern University research confirmed this, finding that models convert sourced claims into asserted facts without signaling the user that the source was lost in transit.</p><h3 class="header" id="h3-93877b6d4798" _key="d33c3d833be3" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-93877b6d4798" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Sycophancy</strong><!--]--></span></h3><p>RLHF training, which is a standard for using human evaluators to fine-tune LLMs, teaches models to prioritize agreement over accuracy. <a href="https://www.nature.com/articles/s41746-025-02008-z"><!--[-->A 2025 study published in <em>npj Digital Medicine</em><!--]--></a> found that sycophantic compliance rates were as high as 100% across five popular LLMs (GPT-4, GPT-4o, GPT-4o-mini, Llama 3-8B, and Llama 3-70B) when given medically illogical prompts.</p><p>Itâs not that the model lacked knowledge, it just found agreement to be the path of least resistance for maximizing its reward function. Sycophancy worsens with scale, and interestingly, <a href="https://arxiv.org/pdf/2602.01002"><!--[-->responds negatively to post-training alignment<!--]--></a>, which often means trying to fix the issue simply makes it worse.</p><h3 class="header" id="h3-76feac9f5324" _key="8e4c3fdf8c4d" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-76feac9f5324" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Scheming during evaluation</strong><!--]--></span></h3><p>Some models behave differently when they detect they're being evaluated, sandbagging on capability tests or quietly pursuing hidden objectives while appearing compliant. <a href="https://arxiv.org/abs/2412.04984"><!--[-->Apollo Research<!--]--></a> documented this across o1, Claude 3.5 Sonnet, Gemini 1.5 Pro, and Llama 3.1 405B in December 2024.</p><p>
1Following the Apollo Research findings, OpenAI's anti-scheming training later reduced deception rates on chat data <a href="https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/"><!--[-->from 31.4% to 14.2%<!--]--></a>. Itâs a tangible improvement, though the researchers cautioned it may be partially explained by models becoming more aware they were being evaluated rather than by genuine alignment.</p><h3 class="header" id="h3-248e8fcd6883" _key="371b02b9d3dc" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-248e8fcd6883" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Why the split matters</strong><!--]--></span></h3><p><a href="https://arxiv.org/html/2604.04788"><!--[-->A 2025 taxonomy paper<!--]--></a> surveying hallucination, sycophancy, sandbagging, and alignment faking confirmed that "mitigations may not transfer across phenomena." When you fix the retrieval pipeline to address hallucination, you have not touched the sycophancy mechanism or scheming behavior at all.</p><p>EMNLP 2025 research also found models that <em>knew</em> the correct answer could still hallucinate a different one and did so with <em>higher</em> certainty than their correct responses. So the confidence signal alone doesn't tell you which failure mode you're dealing with.</p><h2 class="header" id="h2-791eac3f7e0c" _key="df33fe5334da" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h2-791eac3f7e0c" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Failure mode #1: Epistemological mismatch</strong><!--]--></span></h2><p>LLMs draw on parametric knowledge (baked into weights at training) and retrieved knowledge (passed in at inference via RAG). The model has no native mechanism to tag which source a claim came from or to enforce that high-confidence assertions require verification.</p><h3 class="header" id="h3-24b06a041188" _key="9ea8294163d1" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-24b06a041188" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Parallels from journalism</strong><!--]--></span></h3><p>A reporter who confidently publishes a claim without tracing it to a verifiable source is doing exactly what an LLM does when it strips attribution mid-synthesis.</p><p>Thatâs why attribution is a structural output requirement in journalism, not a style preference. Every factual claim links to its origin in the published text. Moreover, the two-source rule adds a corroboration threshold. High-stakes claims can't be stated without independent confirmation from a second source.</p><h3 class="header" id="h3-62fcff0693a3" _key="e4d40b7a779b" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-62fcff0693a3" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Engineering solutions for developers</strong><!--]--></span></h3><p>You fix this by treating attribution as a schema constraint rather than a content preference. That means building provenance tagging into the response object itself, not as a footnote or a "sources consulted" section, but as a structured field attached to each factual claim.</p><p>Claims that can't be linked to a retrieved document get tagged `source: inference` before they reach the user rather than smoothed into confident prose. Microsoft Azure AI Foundry's Grounding with Bing Search already enforces this pattern through its Use and Display Requirements, and Google's NotebookLM takes the same approach with source-linked responses.</p><p>But tagging alone isn't enough if the synthesis step can still override it. That's where assertion gating comes in, a pre-output validation pass that checks each high-confidence claim against a retrieved passage above a defined similarity threshold, downgrading anything that doesn't clear it.</p><p>So before a claim gets stated at high confidence, the system checks whether more than one retrieved document independently supports it. Claims with only a single source get flagged as uncertain rather than asserted as fact. The open-source <a href="https://docs.ragas.io/en/latest/"><!--[-->
1RAG evaluation framework from Exploding Gradients<!--]--></a> (RAGAS) identifies âatomic factual statementsâ and enforces a Faithfulness metric that serves this exact purpose.</p><p><a href="https://arxiv.org/pdf/2509.25494"><!--[-->Northwestern research describes a five-stage pipeline<!--]--></a> (corpus summarization, search planning, parallel thread execution, quality evaluation, and synthesis) with citation chains maintained throughout and unsupported claims rejected rather than passed through. <a href="https://aws.amazon.com/blogs/aws/minimize-ai-hallucinations-and-deliver-up-to-99-verification-accuracy-with-automated-reasoning-checks-now-available/"><!--[-->Amazon Bedrock's Automated Reasoning Checks<!--]--></a> also run formal logic validation against a domain knowledge policy and claims a 99% response accuracy rate doing it, though that's a vendor figure worth pressure-testing in your own evals.</p><h2 class="header" id="h2-29b5c2ef814a" _key="cb9e867477f8" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h2-29b5c2ef814a" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Failure mode #2: Sycophancy</strong><!--]--></span></h2><p>Human evaluators <a href="https://www.anthropic.com/research/towards-understanding-sycophancy-in-language-models"><!--[-->consistently rate agreeable responses higher than corrective ones<!--]--></a>, even when the correction is more accurate. This causes models to learn that agreement is rewarded over accuracy and makes them more likely to offer the answer that someone is looking for rather than the one thatâs based on evidence or facts.</p><h3 class="header" id="h3-9c8cde4dec2d" _key="ff61fcec1c7d" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-9c8cde4dec2d" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Parallels from journalism</strong><!--]--></span></h3><p>There's a phenomenon in media called access journalism. Reporters who cultivate relationships with powerful sources end up softening coverage to preserve that access. The source's approval gets over-weighted relative to the truth of their claims.</p><p>It's a structural issue. The feedback loop distorts the output without even requiring deceptive intentions from anyone involved.</p><p>Thatâs why newsrooms maintain editorial independence. A reporter who cultivates the source is not the person who decides what gets published, while the editor's role remains explicitly adversarial. Newsrooms also enforce a no-pre-approval policy where sources never review conclusions before publication, because if they did, the incentive to please would corrupt the output.</p><h3 class="header" id="h3-cdc326c9f387" _key="b5710c669f32" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-cdc326c9f387" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Engineering solutions for developers</strong><!--]--></span></h3><p>This newsroom solution can also be adjusted to work with sycophantic AI models. A primary model generating a response and the component evaluating that response need different objective functions, otherwise you're asking the sycophant to grade their own work.</p><p>Building an adversarial verification layer (a second model or eval component explicitly permitted to challenge the initial output) addresses this directly. It checks whether responses are based on unverified premises, accept false framing, or suppress contradicting evidence. In fact, <a href="https://www.nature.com/articles/s41746-025-02008-z"><!--[-->npj Digital Medicine published a study<!--]--></a> that found simply giving models explicit "rejection permission" substantially improved performance on illogical requests. Creating a dedicated verification layer formalizes that permission at the architecture level rather than relying on prompts or skills.</p><p><a href="https://aclanthology.org/2025.findings-acl.1141/"><!--[-->An ACL Anthology study<!--]--></a> even outlined a framework called CONSENSAGENT that uses structured prompt optimization between agents to reduce sycophancy in multi-agent debate systems. Sycophancy is often amplified at each step where multiple AI agents are working together. CONSENSAGENT solves this by actively refining prompts at each stage to eliminate unverified assumptions that get passed down to the next stage of the workflow.</p><p>Then thereâs the matter of catching false premises before synthesis even begins. Most sycophantic failures aren't just about the model agreeing with a claim. They're about the model building an entire response on top of something the user assumed to be true that wasn't.</p><p>Premise auditing is a technique that extracts those unverified claims and runs them against the retrieval index before synthesis starts. Premises that fall below a validity threshold trigger a clarification rather than an answer that silently accepts the bad user input. Amazon Bedrock's Automated Reasoning Checks can encode valid domain premises as a formal policy, flagging false user-supplied assumptions before they reach output.</p><p>After that, the final lever is directional. Since the model over-rewards itself for affirmation and under-rewards a challenge, the validation threshold should be optimized to compensate for this design. That means requiring the model to find stronger retrieved evidence to confirm a user's assumptions than to contradict it. Itâs the same logic a reporter applies when deciding how much proof they need to confirm a claim from a potentially biased source.</p><h2 class="header" id="h2-f1589d9aea4d" _key="5ad5fe611bbf" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h2-f1589d9aea4d" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Failure mode #3: Scheming</strong><!--]--></span></h2><p>Models in certain configurations behave differently when they detect they're being evaluated, intentionally underperforming on capability tests or quietly pursuing hidden agendas while appearing compliant. <a href="https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/"><!--[-->The OpenAI/Apollo September 2025 paper<!--]--></a> traced it to "the model's situational awareness of being evaluated."</p><h3 class="header" id="h3-4d04f8615c33" _key="b3154f6d1575" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-4d04f8615c33" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Parallels from journalism</strong><!--]--></span></h3><p>Newsrooms have institutionalized processes for addressing the possibility of deliberate unethical behavior from reporters.</p><p>Two specific mechanisms apply here. First, mandatory process transparency requires reporters to
1show their reasoning to editors even when only conclusions are published, making the reasoning chain auditable internally.</p><p>Second, third-parties with no stake in the result should audit results. Fact-checking teams evaluate work against independent standards, structurally separate from the reporting team. This exists because self-reporting of failures cannot be the primary reliability mechanism in any high-trust information system.</p><h3 class="header" id="h3-1554c2117533" _key="071c3cf03873" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-1554c2117533" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Engineering solutions for developers</strong><!--]--></span></h3><p>Newsrooms require reporters to maintain detailed notes. You keep them not because the reader sees them but because they're the internal record against which conclusions can be audited. Chain-of-thought logging in an AI system serves the same function.</p><p>By retaining model reasoning chains as operational records, which can then be reviewed by teams independent of the production pipeline, you can see when stated reasoning diverges from output behavior.</p><p>OpenAI has already been advocating for "preserving chain-of-thought transparency industry-wide" as a mitigation for model scheming. Braintrust is another project that captures full reasoning traces as first-class operational records with replay capability. This serves not only as a debugging tool but also built-in accountability infrastructure. For agentic deployments, the EU AI Act's General-Purpose AI transparency obligations, in force since August 2025, already require documenting system objectives, capabilities, and intended purposes.</p><p>Sometimes, a model also pretends to be less capable during evaluations so it can avoid triggering stronger safety interventions. But if the model cannot distinguish evaluations from production traffic, it wonât know to underperform on purpose. This can be achieved through blind evaluation sets, where prompts are drawn from real production distributions, rotated regularly, administered through the same interface, and conditioned as production queries. A subject can't behave differently for the camera if they don't know the camera is on.</p><p>Anthropic's open-source Bloom framework does something similar by generating behavioral evaluation scenarios that models haven't specifically trained against, because static benchmarks don't usually catch this kind of adaptive behavior.</p><p>But even good evaluation sets go stale. Behavioral consistency monitoring tracks output distribution across user-facing and internal contexts, flagging any divergence between responses to similar queries under different observable conditions. Braintrust, LangSmith, and Arize Phoenix all support observability pipelines that do exactly this.</p><h2 class="header" id="h2-ac4834c21929" _key="f56ed9a17d21" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h2-ac4834c21929" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>A decision framework for development teams</strong><!--]--></span></h2><p>The table below consolidates the three sections into a reference for sprint planning, not as a summary but as a starting point for prioritization decisions.</p><p><strong>Failure mode</strong></p><p><strong>Causal mechanism</strong></p><p><strong>Journalism parallels</strong></p><p><strong>Engineering solutions</strong></p><p><strong>Implementation layer</strong></p><p><strong>Epistemological mismatch</strong></p><p>No architectural distinction between retrieved knowledge and training-data plausibility. The model produces confident output either way; attributed claims get converted into universal assertions.</p><ul><li>Attribution as structural output requirement.</li><li>Two-source rule as corroboration thresholdâthe claim doesn't go to press at high confidence until independently confirmed.</li><li><strong>Citation-enforced synthesis:</strong> Every claim carries a source document ID as a structured field, not a footnote.</li><li><strong>Assertion gating:</strong> Pre-output validation checks each high-confidence claim against retrieved passages; unsupported claims get downgraded before reaching the user.</li><li><strong>Corroboration thresholds:</strong> `min_sources_for_high_confidence` as a configurable parameter, enforced at the schema level.</li></ul><p>Retrieval + synthesis pipeline</p><p><strong>Sycophancy</strong></p><p>RLHF reward signals over-reward agreement. Human evaluators rate agreeable responses higher than corrective ones even when the correction is more accurate; the model learns the pattern and generalizes it.</p><ul><li>Editorial independence. The reporter who cultivates the source is not the person who decides what gets published.</li><li>No-pre-approval policy. Sources never review conclusions before publication.</li><li><strong>Adversarial verification layer:</strong> A separate model component with its own objective function, explicitly permitted to challenge outputânot self-critique, which replicates the sycophantic dynamic.</li><li><strong>Premise auditing:</strong> Extract the user's implicit factual claims as propositions, validate against the retrieval index before synthesis begins.
1</li><li><strong>Asymmetric confidence thresholds:</strong> Require stronger retrieved evidence to affirm a user's stated belief than to contradict it, compensating for the directional reward bias.</li></ul><p>System architecture + pre-synthesis stage</p><p><strong>Scheming</strong></p><p>Models adjust behavior under detected evaluation context. Behavioral inconsistency is empirically documented; the engineering problem is not resolving the philosophy of intent but designing systems that treat inconsistency as the default assumption.</p><ul><li>Mandatory process transparency. The reasoning chain is auditable internally, independent of the published output.</li><li>Third-party audit by parties structurally separate from the publisher, with no stake in the result.</li><li><strong>CoT logging as audit artifact:</strong> Reasoning traces retained as operational records, reviewable by teams independent of the production pipelineânot just for debugging.</li><li><strong>Blind evaluation sets:</strong> Prompts drawn from real production distributions, rotated regularly, administered so the model cannot distinguish evaluation from production traffic.</li><li><strong>Behavioral consistency monitoring:</strong> Flagging statistically significant divergence between outputs across user-visible and system-internal contexts.</li></ul><p>Evaluation design + operational monitoring</p><p>Not every team needs to address all three failure modes at the same time. Which one to prioritize depends on what you're building and who's using it.</p><h3 class="header" id="h3-d0593def6611" _key="7fb1106f5187" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-d0593def6611" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>RAG-based knowledge tools</strong><!--]--></span></h3><p>If users trust your AI-generated answers without independently verifying them, as they often do in internal documentation tools, research assistants, and knowledge bases, epistemological mismatch is the first problem to solve. Users have no way to detect when a model strips attribution or paraphrases a source into something subtly different.</p><p>A developer reviewing generated code can spot a hallucinated function name. A product manager reading an AI summary of customer feedback usually can't. Provenance tagging and assertion gating need to happen at the system level because they can't happen at the user level.</p><h3 class="header" id="h3-d39c35d76988" _key="35b6f3c6798e" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-d39c35d76988" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Products for users with strong existing beliefs</strong><!--]--></span></h3><p>This covers more ground than it sounds like. Health and wellness apps, financial planning tools, legal research assistants, technical troubleshooting bots. Anywhere users arrive with a strong prior, sycophancy is the dominant risk.</p><p>The <em>npj Digital Medicine</em> study found that GPT-4, GPT-4o, and GPT-4o-mini complied with medically illogical prompts 100% of the time without special instruction otherwise. Users who get validation of a false premise don't just stay misinformed. They leave with more confidence than they arrived with.</p><p>An adversarial verification layer and premise auditing step are worth the added latency in these contexts.</p><h3 class="header" id="h3-21e79544cd23" _key="6a776ec73330" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h3-21e79544cd23" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Agentic systems</strong><!--]--></span></h3><p>In agentic systems where AI models have multi-step autonomy, the threat profile changes entirely. A sycophantic chatbot gives a wrong answer, a sycophantic agent takes a wrong action, and that leads to several more downstream actions built on top of it.</p><p>Scheming compounds this issue. A model that behaves differently under evaluation than in production is one you cannot reliably test. Audit trail infrastru
1cture (CoT logging, blind evaluation sets, behavioral consistency monitoring) has to be in place before deployment, not retrofitted after the first incident.</p><p>Relying on models to self-report their failure isnât good strategy. It's just a hope.</p><h2 class="header" id="h2-ec52c421a7c6" _key="7470d768b4dd" _type="block" data-v-7442e526><a title="Permalink to this heading" href="#h2-ec52c421a7c6" aria-hidden="true" data-v-7442e526></a><span data-v-7442e526><!--[--><strong>Reliability as a design constraint</strong><!--]--></span></h2><p>Newsrooms developed these frameworks because each failure mode destroyed something valuable, with costs severe enough to require structural solutions rather than quietly ignored editorial policies. Epistemological failures hurt credibility. Sycophancy jeopardized editorial independence. Behavioral inconsistency messed with institutional trust.</p><p><a href="https://survey.stackoverflow.co/2025"><!--[-->Stack Overflowâs 2025 Developer Survey<!--]--></a> shows what this looks like by the numbers. Developer trust in AI accuracy fell from 40% to 29% in a single year, even as adoption climbed to 84%. More developers actively distrust AI tool output (46%) than trust it (33%), while only 3% report high trust.</p><p>Experienced developers are the most skeptical, with the highest distrust rate at 20%. That's the phase where credibility is destroyed completely and itâs happening faster than most AI vendors are willing to acknowledge.</p><p>The solutions we described here are available now. What's missing isn't the technical knowledge but the design culture that treats evidence handling as a first-class engineering concern, on par with security and performance.</p><p>You cannot treat factual accuracy as a content moderation problem to address after the first high-profile incident makes it unavoidable. Five hundred years of journalism may have had to learn this the hard way, but you don't.</p><p></p><p></p><p><strong>Sources:</strong></p><ol><li><a href="https://arxiv.org/abs/2509.25498"><!--[--><strong>"Not wrong, but untrue: LLM overconfidence in document-based queries" (Sept. 2025)</strong><!--]--></a></li><li><a href="https://arxiv.org/pdf/2509.25494"><!--[--><strong>"On-premise AI for the newsroom: Evaluating small language models for investigative document search" (Sept. 2025)</strong><!--]--></a></li><li><a href="https://arxiv.org/html/2604.04788"><!--[--><strong>"From hallucination to scheming: A unified taxonomy and benchmark analysis for LLM deception" (2025)</strong><!--]--></a></li><li><a href="https://arxiv.org/abs/2410.22071"><!--[--><strong>"Distinguishing ignorance from error in LLM hallucinations" (2024)</strong><!--]--></a></li><li><a href="https://arxiv.org/pdf/2602.01002"><!--[--><strong>"How RLHF amplifies sycophancy" (Jan. 2026)</strong><!--]--></a></li><li><a href="https://www.nature.com/articles/s41746-025-02008-z"><!--[--><strong>npj Digital Medicineâ"When helpfulness backfires: LLMs and the risk of false medical information due to sycophantic behavior" (Oct. 2025)</strong><!--]--></a></li><li><a href="https://read.dukeupress.edu/critical-ai/article/doi/10.1215/2834703X-11700255/401267/Rethinking-Error-Hallucinations-and"><!--[--><strong>Duke University Pressâ"Rethinking error: Hallucinations and epistemological indifference" (Apr. 2025)</strong><!--]--></a></li><li><a href="https://arxiv.org/abs/2412.04984"><!--[--><strong>Apollo Researchâ"Frontier models are capable of in-context scheming" (Dec. 2024)</strong><!--]--></a></li><li><a href="https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/"><!--[--><strong>OpenAI / Apollo Researchâ"Detecting and reducing scheming in AI models" (Sept. 2025)</strong><!--]--></a></li><li><a href="https://arxiv.org/html/2502.12964v2"><!--[--><strong>EMNLP 2025â"LLMs hallucinate with certainty despite knowing the answer"</strong><!--]--></a></li><li><a href="https://www.pnas.org/doi/10.1073/pnas.2322823121"><!--[--><strong>PNASâ"Fact-checking information from large language models can decrease headline discernment" (Dec. 2024)</strong><!--]--></a></li><li><a href="https://aclanthology.org/2025.findings-acl.1141/"><!--[--><strong>CONSENSAGENT: Towards Efficient and Effective Consensus in Multi-Agent LLM Interactions Through Sycophancy Mitigation - ACL Anthology</strong><!--]--></a></li><li><a href="https://stackoverflow.co/company/press/archive/stack-overflow-2025-developer-survey/"><!--[--><strong>Stack Overflowâs 2025 Developer Survey Reveals Trust in AI at an All Time Low - Press release</strong><!--]--></a></li></ol><!--]--></div><!----></article><aside class="flex--item3 pt12 ml48 lg:ml0"><!----><!----><div class="mb24 pb24 bb bc-black-200"><div class="fc-black-400 tt-uppercase fw-bold fs-body2 mb16"> Author<!--[-->s<!--]--></div><!--[--><a href="/author/ritoban-mukherjee/" class="d-block bar-md p24 my12 ba fc-black h:fc-blue-500 bg-blue-100 bc-transparent" title="View full bio and posts"><div class="d-flex ai-center"><div><div class="fw-bold fs-body3" itemprop="author">
1Ritoban Mukherjee<svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconArrowRight va-middle mtn4 ml2" aria-hidden="true"><path d="m17.52 9.28.56.54-.56.54-7.25 6.93-1.04-1.08 5.97-5.71H2V9h15.24zM15.7 7.5h-2.15L9.23 3.29l1.04-1.08z"></path></svg></div><!----><!----><!----></div><div class="ml-auto ps-relative as-center"><img class="bar-circle bs-sm" width="48" height="48" src="https://cdn.stackoverflow.co/images/jo7n4k8s/production/c5181485849c7a09ec3e175540fa84ac2f76b591-512x512.png?w=96&h=96&auto=format" alt="Ritoban Mukherjee"></div></div><!----></a><!--]--></div><div class="d-flex ai-start gs4 ff-row-wrap" itemprop="keywords"><!----><!--[--><a href="/ai/" class="bc-blue-400 fc-blue-500 h:bg-blue-200 h:bc-blue-500 flex--item ba bar-md fw-bold px8 py2" style="">AI</a><!--]--><!----></div><!----><!----><!--[--><div class="fc-black-400 mt32 tt-uppercase fw-bold fs-body2 mb16"> Recent articles </div><div class="ba bc-black-150 bar-md bar-md pb24 pl24 pr24 bg-black-100"><!--[--><article class="overflow-hidden pt16 mt16" itemscope itemtype="https://schema.org/Article"><div class="d-flex gs32 gsx h100 lg:fd-column"><div class="flex--item12 d-flex fd-column"><time datetime="2026-09-16T17:00:00.000Z" class="flex--item tt-uppercase fw-bold fs-body1 fc-black-500 d-flex mb12" itemprop="datePublished"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconCalendar va-middle mr6 mtn1" aria-hidden="true"><path d="M7.6 2h4.8V.6H14V2h4v16H2V2h4V.6h1.6zM3.5 16.5h13v-8h-13zm4-2H6V13h1.5zm3.25 0h-1.5V13h1.5zm3.25 0h-1.5V13H14zm-6.5-3H6V10h1.5zm3.25 0h-1.5V10h1.5zm3.25 0h-1.5V10H14zM3.5 7h13V3.5H14v1.89h-1.6V3.5H7.6v1.89H6V3.5H3.5z"></path></svg> September 16, 2026</time><a href="/2026/09/16/here-s-what-s-new-on-stack-overflow-for-agents/" class="fc-black h:fc-blue-400"><h1 class="fs-subheading ff-stack-sans-headline lh-sm wmx7 mb0" itemprop="name">From better privacy to our new ChatGPT plugin, here's what's new on Stack Overflow for Agents</h1></a><!----><!----><!----></div><!----></div></article><article class="overflow-hidden pt16 mt16 bt bc-black-200" itemscope itemtype="https://schema.org/Article"><div class="d-flex gs32 gsx h100 lg:fd-column"><div class="flex--item12 d-flex fd-column"><time datetime="2026-09-10T18:01:54.891Z" class="flex--item tt-uppercase fw-bold fs-body1 fc-black-500 d-flex mb12" itemprop="datePublished"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconCalendar va-middle mr6 mtn1" aria-hidden="true"><path d="M7.6 2h4.8V.6H14V2h4v16H2V2h4V.6h1.6zM3.5 16.5h13v-8h-13zm4-2H6V13h1.5zm3.25 0h-1.5V13h1.5zm3.25 0h-1.5V13H14zm-6.5-3H6V10h1.5zm3.25 0h-1.5V10h1.5zm3.25 0h-1.5V10H14zM3.5 7h13V3.5H14v1.89h-1.6V3.5H7.6v1.89H6V3.5H3.5z"></path></svg> September 10, 2026</time><a href="/2026/09/10/re-introducing-developer-story/" class="fc-black h:fc-blue-400"><h1 class="fs-subheading ff-stack-sans-headline lh-sm wmx7 mb0" itemprop="name">(Re)introducing Developer Story</h1></a><!----><!----><!----></div><!----></div></article><article class="overflow-hidden pt16 mt16 bt bc-black-200" itemscope itemtype="https://schema.org/Article"><div class="d-flex gs32 gsx h100 lg:fd-column"><div class="flex--item12 d-flex fd-column"><time datetime="2026-09-03T16:00:00.000Z" class="flex--item tt-uppercase fw-bold fs-body1 fc-black-500 d-flex mb12" itemprop="datePublished"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconCalendar va-middle mr6 mtn1" aria-hidden="true"><path d="M7.6 2h4.8V.6H14V2h4v16H2V2h4V.6h1.6zM3.5 16.5h13v-8h-13zm4-2H6V13h1.5zm3.25 0h-1.5V13h1.5zm3.25 0h-1.5V13H14zm-6.5-3H6V10h1.5zm3.25 0h-1.5V10h1.5zm3.25 0h-1.5V10H14zM3.5 7h13V3.5H14v1.89h-1.6V3.5H7.6v1.89H6V3.5H3.5z"></path></svg> September 3, 2026</time><a href="/2026/09/03/security-control-and-accessibility-si-2026-6/" class="fc-black h:fc-blue-400"><h1 class="fs-subheading ff-stack-sans-headline lh-sm wmx7 mb0" itemprop="name">Elevating security, control, and accessibility: Stack Internal 2026.6</h1></a><!----><!----><!----></div><!----></div></article><article class="overflow-hidden pt16 mt16 bt bc-black-200" itemscope itemtype="https://schema.org/Article"><div class="d-flex gs32 gsx h100 lg:fd-column"><div class="flex--item12 d-flex fd-column"><time datetime="2026-08-24T14:00:00.000Z" class="flex--item tt-uppercase fw-bold fs-body1 fc-black-500 d-flex mb12" itemprop="datePublished"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconCalendar va-middle mr6 mtn1" aria-hidden="true"><path d="M7.6 2h4.8V.6H14V2h4v16H2V2h4V.6h1.6zM3.5 16.5h13v-8h-13zm4-2H6V13h1.5zm3.25 0h-1.5V13h1.5zm3.25 0h-1.5V13H14zm-6.5-3H6V10h1.5zm3.25 0h-1.5V10h1.5zm3.25 0h-1.5V10H14zM3.5 7h13V3.5H14v1.89h-1.6V3.5H7.6v1.89H6V3.5H3.5z"></path></svg> August 24, 2026</time>
1<a href="/2026/08/24/responsible-ai-adoption-needs-developer-workflow-design/" class="fc-black h:fc-blue-400"><h1 class="fs-subheading ff-stack-sans-headline lh-sm wmx7 mb0" itemprop="name">Responsible AI adoption needs developer workflow design</h1></a><!----><!----><!----></div><!----></div></article><!--]--></div><div class="fc-black-400 mt32 tt-uppercase fw-bold fs-body2 mb16"> Latest Podcast </div><!--[--><article class="overflow-hidden p24 bg-black-600 fc-white ba bc-transparent bar-md mb16" itemscope itemtype="https://schema.org/Article"><div class="d-flex gs32 gsx h100 lg:fd-column"><div class="flex--item12 d-flex fd-column"><time datetime="2026-09-23T04:40:00.000Z" class="flex--item tt-uppercase fw-bold fs-body1 fc-black-300 d-flex mb12" itemprop="datePublished"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconCalendar va-middle mr6 mtn1" aria-hidden="true"><path d="M7.6 2h4.8V.6H14V2h4v16H2V2h4V.6h1.6zM3.5 16.5h13v-8h-13zm4-2H6V13h1.5zm3.25 0h-1.5V13h1.5zm3.25 0h-1.5V13H14zm-6.5-3H6V10h1.5zm3.25 0h-1.5V10h1.5zm3.25 0h-1.5V10H14zM3.5 7h13V3.5H14v1.89h-1.6V3.5H7.6v1.89H6V3.5H3.5z"></path></svg> September 23, 2026</time><a href="/2026/09/23/multiplayer-ai-why-your-team-and-its-agents-need-a-group-chat/" class="fc-white h:fc-blue-400"><h1 class="fs-subheading d:fc-white ff-stack-sans-headline lh-sm wmx7 mb0" itemprop="name">Multiplayer AI: Why your team (and its agents) need a group chat</h1></a><!----><!----><!----></div><!----></div></article><!--]--><!--]--></aside></div></div><section class="bg-black-50 wmx7 mx-auto sm:pl24 sm:pr24 py64 sm:pt32 sm:pb32 ff-sans" id="comments"><!--[--><h2>Add to the discussion</h2><div class="p24 bs-md bar-md d-flex"><div class="m-auto d-inline-flex sm:fd-column ai-center ta-center"><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconGlyph mr12 sm:mb12" aria-hidden="true"><path d="M14.83 10.7a10 10 0 0 0-1.31 2.21l-.02.03q-.5 1.17-.68 2.46v.02a10 10 0 0 0-.03 2.58H3v-2.6h8.5l.06-.32-8.2-2.17.69-2.5 8.29 2.19.12-.25L5 8.1l1.32-2.26 7.55 4.3.19-.2-6.2-6.11L9.75 2l6.29 6.2.66.66a10 10 0 0 0-1.86 1.84"></path></svg><div>Login with your <strong>stackoverflow.com</strong> account to take part in the discussion.</div></div></div><!--]--><!----><!----><!--[--><!--]--></section><!----><!--]--><!--]--><!--[--><footer id="footer-global" data-v-df5a0bb6><div class="footer__inner p24 pb0 m12 bg-black-150 fs-body2" data-v-df5a0bb6><div class="d-flex sm:fd-column" data-v-df5a0bb6><section class="flex--item4" data-v-df5a0bb6><svg width="32" height="32" viewBox="0 0 32 32" class="svg-icon IconGlyph32 mb24" aria-hidden="true" title="Stack Overflow" data-v-df5a0bb6><path d="M23.8 17.22a18 18 0 0 0-2.2 3.87l-.02.06q-.83 2.04-1.15 4.3v.03A18 18 0 0 0 20.4 30H4v-4.55h14.23l.1-.56L4.6 21.1l1.14-4.4 13.88 3.85.2-.44-12.46-7.43 2.2-3.94 12.65 7.54.3-.37L12.17 5.2 15.28 2l10.53 10.86 1.1 1.15q-1.75 1.4-3.11 3.21"></path></svg><!----></section><div class="sections flex--item8 d-flex md:mb32 md:mt24" data-v-df5a0bb6><div class="flex--item6 d-flex md:fd-column" data-v-df5a0bb6><section class="flex--item6" data-v-df5a0bb6><h3 id="footer-products-heading" data-v-df5a0bb6> Our Stack </h3><nav aria-labelledby="footer-products-heading" data-v-df5a0bb6><ul class="list-reset" data-v-df5a0bb6><li data-v-df5a0bb6><a href="https://stackoverflow.co/internal/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Stack Internal</a><ul class="list-reset" data-v-df5a0bb6><li data-v-df5a0bb6><a href="https://stackoverflow.co/internal/features/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" class="sub-link" data-v-df5a0bb6>Features</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.co/internal/customers/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" class="sub-link" data-v-df5a0bb6>Customers</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.co/internal/uses/ai?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" class="sub-link" data-v-df5a0bb6>Use cases</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.co/internal/security/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" class="sub-link" data-v-df5a0bb6>Security</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.co/internal/integrations/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" class="sub-link" data-v-df5a0bb6>Integrations</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.co/internal/pricing/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" class="sub-link" data-v-df5a0bb6>Pricing</a></li></ul></li><li class="mt12" data-v-df5a0bb6><a href="https://stackoverflow.co/data-licensing/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Stack Data Licensing</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.co/advertising/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Stack Ads</a></li><li class="mt12" data-v-df5a0bb6><a href="https://stackoverflow.co/partnerships/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Partnerships</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.co/internal/services/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Services</a></li><li class="mt12" data-v-df5a0bb6><a href="https://stackoverflow.com/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Stack Overflow</a></li></ul></nav></section><section class="flex--item6" data-v-df5a0bb6><h3 id="footer-company-heading" class="md:mt32" data-v-df5a0bb6> Company </h3><nav aria-labelledby="footer-company-heading" data-v-df5a0bb6><ul class="list-reset" data-v-df5a0bb6><li data-v-df5a0bb6><a href="https://stackoverflow.co/company/leadership/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Leadership</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.co/company/press/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Press</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.co/company/careers/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Careers</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.co/company/social-impact/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Social Impact</a></li></ul></nav></section></div><div class="flex--item6 d-flex md:fd-column" data-v-df5a0bb6><section class="flex--item6" data-v-df5a0bb6><h3 id="footer-support-heading" class="md:mt32" data-v-df5a0bb6> Support </h3><nav aria-labelledby="footer-support-heading" data-v-df5a0bb6><ul class="list-reset" data-v-df5a0bb6><li data-v-df5a0bb6><a href="https://stackoverflow.com/contact/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Contact</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.com/help/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Stack Overflow help</a></li><li data-v-df5a0bb6><a href="https://support.stackenterprise.co/support/solutions/22000023964" rel="noopener noreferrer" data-v-df5a0bb6>Stack Internal help</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.co/internal/implement/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Implement Stack Internal</a></li><li class="mt12" data-v-df5a0bb6><a href="https://policies.stackoverflow.co/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Terms</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.com/legal/privacy-policy/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Privacy policy</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.com/legal/cookie-policy/?utm_source=blog&utm_medium=referral&utm_campaign=footer" rel="noopener noreferrer" data-v-df5a0bb6>Cookie policy</a></li><li class="mt12" data-v-df5a0bb6><button class="flex--item mr24 c-pointer s-btn s-btn__unset p0 as-start fc-black h:fc-blue-500" type="button" title="Open cookie consent settings" data-v-df5a0bb6>
1 Your Privacy Choices <svg width="34" height="16" viewBox="0 0 34 16" class="svg-icon IconServiceCCPA va-middle mtn2 ml4 native" aria-hidden="true" data-v-df5a0bb6><path fill="#06f" d="M25.2 0a8 8 0 0 1 7.94 8c0 4.46-3.52 8-7.94 8H7.95A7.9 7.9 0 0 1 0 8c0-4.46 3.52-8 7.95-8zM7.95 1.37A6.6 6.6 0 0 0 1.36 8a6.6 6.6 0 0 0 6.59 6.63h7.71l3.52-13.26zm19.52 3.2a.7.7 0 0 0-.91 0l-2.5 2.52-2.5-2.52a.7.7 0 0 0-.9 0 .7.7 0 0 0 0 .92L23.16 8l-2.39 2.51a.7.7 0 0 0 0 .92c.23.23.68.23.9 0l2.5-2.52 2.5 2.52c.23.23.68.23.91 0a.7.7 0 0 0 0-.92L25.08 8l2.39-2.51a.7.7 0 0 0 0-.92m-13.98.12c.22-.23.68-.23 1.02 0 .23.22.35.68.12.91l-4.8 5.6c-.12.11-.23.23-.34.23a.7.7 0 0 1-.8-.12L6.17 8.8a.7.7 0 0 1 0-.91.7.7 0 0 1 .92 0l2.05 1.94z"></path></svg></button></li></ul></nav></section><section class="flex--item6" data-v-df5a0bb6><h3 id="footer-content-heading" class="md:mt32" data-v-df5a0bb6> Elsewhere </h3><nav aria-labelledby="footer-content-heading" data-v-df5a0bb6><ul class="list-reset" data-v-df5a0bb6><li data-v-df5a0bb6><a href="https://stackoverflow.blog/" rel="noopener noreferrer" data-v-df5a0bb6>Blog</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.blog/newsletter/" rel="noopener noreferrer" data-v-df5a0bb6>Dev Newsletter</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.blog/podcast/" rel="noopener noreferrer" data-v-df5a0bb6>Podcast</a></li><li data-v-df5a0bb6><a href="https://stackoverflow.blog/releases/" rel="noopener noreferrer" data-v-df5a0bb6>Releases</a></li><li data-v-df5a0bb6><a href="https://survey.stackoverflow.co/" rel="noopener noreferrer" title="Annual Developer Survey" data-v-df5a0bb6>Dev Survey</a></li></ul></nav></section></div></div></div><svg width="175" height="22" viewBox="0 0 175 19" class="svg-icon IconLogo flex--item12 w100 h-auto fc-black-050 ps-relative" aria-hidden="true" title="Stack Overflow" data-v-df5a0bb6><g fill="currentColor"><path d="M18.52 9.69a13.7 13.7 0 0 0-4.3 5.47l-.01.05a13 13 0 0 0-.93 3.32v.02q-.13.92-.13 1.88c0 .55.07 1.04.13 1.57H0v-3.47h11.5l.07-.43L.5 15.16l.93-3.4 11.21 2.98.16-.34L2.72 8.66 4.5 5.6l10.22 5.83.25-.29L6.6 2.89 9.12.4zM28.9 5.78c3.3 0 5.98 2.15 5.98 5.24h-2.96c-.13-1.67-1.09-2.66-3.11-2.66-2.03 0-3.04.81-3.04 1.95 0 3.42 9.34.33 9.34 6.58 0 2.89-2.28 4.51-6.13 4.51-3.57 0-6.22-2.25-6.33-5.27h3c.14 1.7 1.59 2.71 3.48 2.71 1.78 0 3.07-.68 3.07-1.95 0-3.14-9.37-.28-9.37-6.58 0-2.45 2.33-4.53 6.08-4.53M49.06 8.36c1.75 0 2.96.7 3.75 1.95V8.7h2.76V21H52.8v-1.53c-.74 1.2-1.93 1.93-3.75 1.93-3.44 0-5.55-2.69-5.55-6.5 0-3.83 2.1-6.54 5.55-6.54m.43 2.48c-2.02 0-3.14 1.7-3.14 4.05 0 2.33 1.12 4 3.14 4 2.23 0 3.32-1.72 3.32-4s-1.1-4.05-3.32-4.05M63.26 8.36c3.37 0 5.73 2.48 5.75 5.2h-2.89c-.07-1.35-.88-2.7-2.86-2.7-2.4 0-3.44 1.75-3.44 4.03s1.04 3.98 3.44 3.98c2.06 0 2.87-1.45 2.94-3h2.86c-.05 3.2-2.3 5.53-5.8 5.53-3.8 0-6.27-2.69-6.27-6.5 0-3.83 2.48-6.54 6.27-6.54M91.52 5.78c4.91 0 7.6 3.57 7.6 7.8 0 4.22-2.71 7.82-7.6 7.82s-7.62-3.55-7.62-7.82c0-4.23 2.7-7.8 7.62-7.8m0 2.58c-3.22 0-4.68 2.68-4.68 5.22s1.44 5.24 4.68 5.24 4.6-2.71 4.6-5.24-1.38-5.22-4.6-5.22M117.45 8.36c3.34 0 6.2 2.7 5.6 7.44h-9c.23 1.8 1.3 3.1 3.4 3.1 1.44 0 2.35-.89 2.73-1.9h2.89a5.5 5.5 0 0 1-5.62 4.4c-3.83 0-6.3-2.69-6.3-6.5 0-3.83 2.47-6.54 6.3-6.54m0 2.48c-1.93 0-2.97 1.11-3.35 2.68h6.23c-.15-1.46-1.06-2.68-2.88-2.68M150.87 8.36c3.8 0 6.28 2.7 6.28 6.53 0 3.82-2.49 6.5-6.28 6.5s-6.3-2.68-6.3-6.5 2.47-6.53 6.3-6.53m0 2.48c-2.3 0-3.47 1.77-3.47 4.05 0 2.33 1.16 4 3.47 4 2.3 0 3.47-1.67 3.47-4s-1.17-4.05-3.47-4.05M40.5 8.71h2.1v2.36h-2.1v7.62h2.46V21h-5.22v-9.93H36V8.7h1.74V6h2.76zM73.19 13.65l3.97-4.94h3.27l-4 4.97 4.4 7.32h-3.08l-2.99-5.3-1.57 1.98V21H70.4V6h2.79zM105.34 18.56l2.79-9.85h2.96L107.24 21h-3.77L99.62 8.71h2.96zM131.22 8.71V11h-3.97v10h-2.76V8.71zM139.2 4.54v2.35h-1.78q-1.18.01-1.19 1.09v.73h2.71V11h-2.7v10h-2.77V11h-1.26V8.71h1.26v-.73c0-2.2 1.27-3.44 3.93-3.44zM143.17 4.54V21h-2.78V4.54zM162.14 17.96l2.05-9.25h3l2.06 9.27 1.82-9.27h2.94L171.05 21h-3.67l-1.67-7.5-1.67 7.5h-3.67L157.4 8.71h2.96z"></path></g></svg></div><div class="px24 pb24 d-flex ai-center ff-row-wrap sm:fd-column sm:ta-center" data-v-df5a0bb6><div class="fs-caption fc-black-400 mr-auto sm:mr0 sm:mb12" data-v-df5a0bb6> Site design / logo © 2026 Stack Exchange Inc. </div><div class="d-flex gs8 ai-center flex--item mr12 sm:mr0" data-v-df5a0bb6><div class="flex--item s-toggle-switch s-toggle-switch__multiple"><!--[--><!--[--><input value="theme-light" type="radio" name="theme" id="theme-light"><label for="theme-light">Light</label><!--]--><!--[--><input value="theme-dark" type="radio" name="theme" id="theme-dark"><label for="theme-dark">Dark</label><!--]--><!--[--><input value="theme-system" type="radio" name="theme" id="theme-system"><label for="theme-system">Auto</label><!--]--><!--]--></div></div><div class="d-flex ai-start sm:mt12" data-v-df5a0bb6><a class="flex--item fc-black p6 h:bg-orange h:fc-white" href="https://linkedin.com/company/stack-overflow/" title="LinkedIn" rel="external" data-v-df5a0bb6><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconServiceLinkedIn" aria-hidden="true" data-v-df5a0bb6><path fill="#0966c2" d="M16.22 2C17.2 2 18 2.8 18 3.78v12.44c0 .98-.8 1.78-1.78 1.78H3.78C2.8 18 2 17.2 2 16.22V3.78C2 2.8 2.8 2 3.78 2zM4.45 15.78h2.39v-7.7H4.45zm8.52-7.94a2.7 2.7 0 0 0-2.32 1.27V8.07H8.36v7.7h2.3v-4.04c0-1.08.49-1.72 1.44-1.72.88 0 1.3.61 1.3 1.72v4.05h2.38V10.9
1c0-2.06-1.17-3.06-2.8-3.06M5.63 4.22c-.78 0-1.4.64-1.4 1.42 0 .79.62 1.43 1.4 1.43s1.41-.64 1.41-1.43-.63-1.42-1.4-1.42"></path></svg></a><a class="flex--item fc-black p6 h:bg-orange h:fc-white" href="https://x.com/stackoverflow/" title="Twitter/X" rel="external" data-v-df5a0bb6><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconServiceX" aria-hidden="true" data-v-df5a0bb6><path d="M14.6 2.77h2.45L11.7 8.89 18 17.23h-4.94L9.2 12.18l-4.42 5.05H2.31l5.74-6.55L2 2.77h5.06l3.5 4.62zm-.86 13h1.36L6.32 4.15H4.86z"></path></svg></a><a class="flex--item fc-black p6 h:bg-orange h:fc-white" href="https://www.threads.net/@thestackoverflow" title="Threads" rel="external" data-v-df5a0bb6><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconServiceThreads" aria-hidden="true" data-v-df5a0bb6><path d="m14.4 9.38-.21-.1c-.13-2.4-1.44-3.77-3.64-3.78h-.03c-1.32 0-2.41.56-3.08 1.58l1.2.83a2.1 2.1 0 0 1 1.9-.92q1.09 0 1.62.62.39.46.52 1.24a9 9 0 0 0-2.08-.1c-2.1.12-3.44 1.34-3.35 3.03.05.86.48 1.6 1.21 2.08.62.41 1.42.61 2.24.57a3.3 3.3 0 0 0 2.55-1.24 4.6 4.6 0 0 0 .87-2.28q.8.47 1.12 1.23c.37.84.39 2.23-.75 3.37-1 .99-2.19 1.42-3.99 1.43q-3.01-.03-4.5-1.9Q4.62 13.25 4.58 10 4.61 6.74 6 4.97q1.48-1.88 4.5-1.91 3.04.02 4.58 1.92.75.93 1.13 2.3l1.42-.38a8 8 0 0 0-1.43-2.87c-1.3-1.61-3.23-2.44-5.7-2.46q-3.72.04-5.64 2.47-1.7 2.17-1.74 5.95V10q.03 3.79 1.74 5.96 1.92 2.44 5.63 2.47h.01c2.2-.02 3.74-.6 5.01-1.86a4.6 4.6 0 0 0 1.07-5.03 4.3 4.3 0 0 0-2.17-2.16m-3.77 3.56c-.92.05-1.87-.36-1.92-1.24-.04-.65.46-1.38 1.97-1.47a9 9 0 0 1 2.03.14c-.17 2.17-1.19 2.52-2.08 2.57"></path></svg></a><a class="flex--item fc-black p6 h:bg-orange h:fc-white" href="https://www.instagram.com/thestackoverflow/" title="Instagram" rel="external" data-v-df5a0bb6><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconServiceInstagram" aria-hidden="true" data-v-df5a0bb6><path d="M10 3.44c2.14 0 2.39 0 3.23.05.78.03 1.2.16 1.49.27.37.15.64.32.92.6s.45.55.6.92c.1.28.24.7.27 1.49.04.84.05 1.1.05 3.23 0 2.14 0 2.39-.05 3.23-.03.78-.16 1.2-.27 1.49-.15.37-.32.64-.6.92s-.55.45-.92.6c-.28.1-.7.24-1.49.27-.84.04-1.1.05-3.23.05-2.14 0-2.39-.01-3.23-.05a4 4 0 0 1-1.49-.27 2.5 2.5 0 0 1-.92-.6 2.5 2.5 0 0 1-.6-.92c-.1-.28-.24-.7-.27-1.49-.04-.84-.05-1.1-.05-3.23 0-2.14.01-2.39.05-3.23.03-.78.16-1.2.27-1.49.15-.37.32-.64.6-.92s.55-.45.92-.6c.28-.1.7-.24 1.49-.27.84-.04 1.1-.05 3.23-.05M10 2c-2.17 0-2.45 0-3.3.05-.85.04-1.43.17-1.94.37-.53.2-.97.48-1.42.92s-.72.9-.92 1.42c-.2.5-.33 1.09-.37 1.94S2 7.83 2 10s0 2.45.05 3.3c.04.85.17 1.43.37 1.94.2.53.48.97.92 1.42s.9.72 1.42.92c.5.2 1.09.33 1.94.37S7.83 18 10 18s2.45 0 3.3-.05a6 6 0 0 0 1.94-.37c.53-.2.97-.48 1.42-.92s.72-.9.92-1.42c.2-.5.33-1.09.37-1.94S18 12.17 18 10s0-2.45-.05-3.3a6 6 0 0 0-.37-1.94c-.2-.53-.48-.97-.92-1.42s-.9-.72-1.42-.92c-.5-.2-1.09-.33-1.94-.37S12.17 2 10 2m0 3.9a4.1 4.1 0 1 0 0 8.2 4.1 4.1 0 0 0 0-8.2m0 6.77a2.67 2.67 0 1 1 0-5.34 2.67 2.67 0 0 1 0 5.34m4.27-7.9a.96.96 0 1 0 0 1.92.96.96 0 0 0 0-1.92"></path></svg></a><a class="flex--item fc-black p6 h:bg-orange h:fc-white" href="https://www.youtube.com/c/StackOverflowOfficial" title="YouTube" rel="external" data-v-df5a0bb6><svg width="20" height="20" viewBox="0 0 20 20" class="svg-icon IconServiceYouTube" aria-hidden="true" data-v-df5a0bb6><path fill="#ff0034" d="M10 3.07c-8.85 0-9 .79-9 6.93s.16 6.93 9 6.93 9-.79 9-6.93-.16-6.93-9-6.93m2.88 7.23L8.84 12.2c-.35.16-.64-.02-.64-.41V8.22c0-.39.29-.57.64-.4l4.04 1.88c.36.17.36.44 0 .6"></path></svg></a></div></div></footer><!----><!--]--><!--]--></div><div id="teleports"></div>
1<script>window.__NUXT__={};window.__NUXT__.config={public:{branch:"main",baseUrl:"https://stackoverflow.blog",context:"production",googleAnalytics:"G-DTTX7JCXEJ",onetrust:"d06435e0-17fb-4659-9a14-1930a6e0be80-test",onetrustRulesetId:"",onetrustCategories:{necessary:"C0001",performance:"C0002",functional:"C0003",targeting:"C0004"},googleMaps:"AIzaSyCFwzoPZ5dOx677jFutgxPd8J2ts1MyPho",googleRecaptchaSiteKey:"6LfCpfEjAAAAAFGHv9bD40vUeqx7Yn8GaacEVP2D",stackappClientId:25117,stackappOauthUrl:"",stackappKey:"GVgf4GPoPeC3eco9n8AEPg((",provenanceCookie:"stackex_provenance",provenanceSessionCookie:"stackex_provenance_session",sanity:{additionalClients:{},apiVersion:"2026-05-27",dataset:"production",disableSmartCdn:false,perspective:"raw",projectId:"jo7n4k8s",queryEndpoint:"",stega:{enabled:true,studioUrl:"https://cms.stackoverflow.co"},token:"",useCdn:true,withCredentials:false,visualEditing:{keepStegaOnCopy:false,mode:"live-visual-editing",previewMode:{enable:"/api/preview/enable",disable:"/api/preview/disable"},previewModeId:"",proxyEndpoint:"/_sanity/visual-editing/fetch",studioUrl:"https://cms.stackoverflow.co",token:"",zIndex:""}},algolia:{apiKey:"170e91ef17b91a8afe6411a133af1309",applicationId:"GGZPPCDRMJ",lite:true,cache:false,instantSearch:false,docSearch:{},recommend:"",globalIndex:"",useFetch:false}},app:{baseURL:"/",buildId:"4900ffdd-7672-4b60-9bc4-4d8cbbb19c70",buildAssetsDir:"/_nuxt/",cdnURL:""}}</script>
1<script type="application/json" data-nuxt-data="nuxt-app" data-ssr="true" id="__NUXT_DATA__" data-src="/2026/06/08/what-can-500-years-of-journalism-teach-developers-about-ai-trustworthiness/_payload.json?_b=4900ffdd-7672-4b60-9bc4-4d8cbbb19c70">[{"state":1,"once":5,"_errors":6,"serverRendered":8,"path":9,"prerenderedAt":-1},["Reactive",2],{"$ssanity-preview":3,"$sarticleModal":3,"$ssanity-preview-environment":4},false,"checking",["Set"],["ShallowReactive",7],{"sanity-kshdV7g2eEekPe3B1ynQWz7Kp9gOcWwFl1uESeZJDUo":-1,"sanity-maVjtXAmyqa5LClU1dW-VgM6oQb3JirXEvi78C0SoAg":-1},true,"\u002F2026\u002F06\u002F08\u002Fwhat-can-500-years-of-journalism-teach-developers-about-ai-trustworthiness\u002F"]</script>
1</body></html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.