A model can produce plausible claims that are unsupported or conflict with evidence.
<div class="scene"><div class="heading">CLAIM CHECK</div><div class="paper">근거 문서 <b>배포일: 9월 1일</b></div><div class="claim">모델 답변 <b id="claim-date">9월 1일</b></div><div class="result" id="claim-result">근거와 일치</div></div>.scene{width:min(94vw,800px);height:min(88vh,326px);padding:clamp(10px,2.7vmin,19px);border:1px solid var(--line);border-radius:13px;background:var(--surface);font:500 clamp(13px,2.4vmin,17px)/1.3 var(--font-sans,sans-serif);position:relative;overflow:hidden}.scene .mono{font-family:ui-monospace,SFMono-Regular,monospace}.scene .muted{color:var(--muted)}.scene .accent{color:var(--accent)}.scene .heading{font-weight:750;color:var(--accent);margin-bottom:clamp(5px,1.6vmin,12px)}.paper,.claim{display:flex;justify-content:space-between;gap:8px;padding:clamp(8px,2.5vmin,18px);border:1px solid var(--line);border-radius:7px;margin:7px 0;background:var(--bg)}.claim b{color:var(--accent)}.claim.bad b{color:#c05545}.result{border-left:4px solid var(--accent);padding:8px;background:var(--bg)}.result.bad{border-color:#c05545;color:#c05545}let bad=false;function draw(){bad=!bad;document.getElementById('claim-date').textContent=bad?'9월 9일':'9월 1일';document.querySelector('.claim').classList.toggle('bad',bad);document.getElementById('claim-result').textContent=bad?'불일치: 근거에 없는 날짜':'근거와 일치';document.getElementById('claim-result').classList.toggle('bad',bad)}draw();setInterval(draw,1500)A generative model composes likely text, so fluency is not proof of truth. It may misread supplied evidence or fill gaps with unsupported claims.
The demo compares a date in evidence with a date in an answer and marks the mismatch. String matching cannot detect every factual error; important claims need independent checks.
Account for it when reviewing model answers or defining grounded-answer quality.