A report went through six AI handoffs. Most of its facts did not arrive intact.
Chinese whispers, played by AI. Newsrooms and publishers now pass the same material through chains of AI steps: an analyst note or press release becomes a summary, an article, a newsletter item, a social post. Each step only sees the one before it, and nobody compares the end with the start. We did, fact by fact, on six models from three vendors.
1 · The decay
Each handoff loses a little. Six lose most of it.
Share of original facts in each state after every step, averaged over all chains (three documents, six models, three runs each).
Intact: present and correctDistorted: changed, or only partly keptMissing
2 · One report, three handoffs
Watch one analyst note change
A real chain from the experiment (GPT-4.1 mini), shown at the start, after the first handoff, after the third and after the sixth. Green is a fact that is still correct. Red is a fact that was changed or a claim that was added. Facts that have disappeared are listed under each step, struck through.
Original note11 of 11
Northgate Precision Industries reported full-year revenue of $1.9 billion, down 6%. Operating margin fell to 4.1% from 7.8%. Net debt stands at 4.6 times EBITDA. Free cash flow was negative $85 million and the dividend has been suspended. Its largest customer, which accounts for 28% of revenue, has not yet renewed its contract. Management guides to revenue flat to down 3% next year. We rate the stock SELL with a 12-month price target of $14, against a current price of $19. Our view could change if the contract is renewed, which we think has roughly a 50% chance. This note is based on unaudited preliminary figures.
Step 1 · Manager summary10 of 11
Northgate Precision reported FY revenue of $1.9B, down 6%, with operating margin falling to 4.1% from 7.8%. Net debt is 4.6x EBITDA; free cash flow was negative $85M. Dividend suspended. Its largest customer (28% revenue) has not renewed its contract1. Management expects flat to 3% lower revenue next year. We rate the stock SELL with a $14 target (current $19), pending contract renewal, which has a ~50% chance. Figures are unaudited preliminary.
1 "Not yet renewed" lost its "yet". An open question starts to read as settled.
Step 3 · Newsletter bullets6 of 11
• Northgate Precision’s FY revenue declined 6% to $1.9B, with operating margins dropping from 7.8% to 4.1%, and free cash flow turning negative $85M. • The largest customer (28% of revenue) did not renew its contract1, creating significant risk; management expects flat to 3% revenue decline next year. • Dividend suspended; stock downgraded2 to SELL with a $14 target price3due to financial challenges and a 50% chance of contract renewal4.
1 Now stated as a decision the customer made.
2 Nothing was downgraded. The note simply rated it SELL.
3 The current price of $19 is gone, so the target no longer says how far the stock could fall.
4 The 50% chance was a reason the view could change. Now it is a reason for the SELL.
Missing by nowNet debt 4.6 times EBITDA Figures are unaudited and preliminary
Step 6 · Client update4 of 11
Northgate Precision reported a 6% revenue decline to $1.9 billion, with operating margins falling to 4.1%1. Free cash flow turned negative $85 million this period. Additionally, their largest customer, accounting for 28% of revenue, will not be renewing contracts2, which is expected to contribute to up to a 3% revenue decline next year3. As a result4, dividends have been suspended, and the stock has been downgraded5 to SELL with a target price of $146. We will continue to monitor the situation closely and keep you updated on any further developments.
1 The comparison with 7.8% is gone, so the size of the fall is lost.
2 "Not yet renewed" has become "will not be renewing".
3 Management's guidance is now blamed on the lost customer. The note made no such link.
4 An invented cause: the note never said why the dividend was suspended.
5 Still a downgrade nobody made.
6 Still no current price to compare with.
Missing by nowNet debt 4.6 times EBITDA View could change if the contract is renewed (about 50% chance) Figures are unaudited and preliminary
Highlights were marked by hand and checked against the AI judge's verdict for each step. Scores are the judge's count of facts still intact. Step through every model and document yourself in section 6.
3 · What dies first
The warnings go first. The verdict stays, but often bent.
We tagged every original fact as a number, a decision (a rating or recommendation), an uncertainty (a condition or an "only if"), or a caveat (a limit on how far to trust the result). This is where each kind stood after six handoffs.
Most fragile facts
Most durable facts
4 · Across all six models
It was not one model. Every vendor changed meaning, not only detail.
Real outputs from the experiment, quoted exactly. Each card shows the original and what the final message said after six handoffs.
An open question became a decision
Original: "Its largest customer, which accounts for 28% of revenue, has not yet renewed its contract."
"…their largest customer, accounting for 28% of revenue, will not be renewing contracts."
GPT-4.1 mini
Then it became the cause
Original: the same sentence. The note never said a customer was lost.
"This underperformance was primarily due to the loss of a key customer."
Gemini 2.5 Flash
A share price became an old target
Original: "We rate the stock SELL with a 12-month price target of $14, against a current price of $19."
"…a recent analyst downgrade of Northgate Precision to SELL with a $14 price target, down from $19."
Claude Haiku 4.5. Nothing was downgraded, and $19 was the share price.
Two facts were joined by an invented "because"
Original: the dividend was suspended. Separately, a customer had not yet renewed.
"…their largest customer … has not renewed their contract, prompting management to suspend dividend payments."
Claude Sonnet 4.5
Someone else's trial became "ours"
Original: "The trial was funded by the manufacturer. The authors say a larger phase 3 trial is needed before the drug can be recommended."
"Our phase 2 trial of zentrolimab has shown promising results … our research team recommends proceeding to a larger phase 3 trial."
Claude Sonnet 4.5. A caution became a go-ahead.
A survey answer became real behaviour
Original: "Fees remained the top stated reason for switching (41%)." Separately, only 18% had actually switched.
"Only 18% have switched banks in the past five years, with fees driving 41% of those moves."
Claude Sonnet 4.5
Advice nobody gave
Original: an analyst note. It gave a rating and no instruction to any client.
"We recommend reviewing your position in light of these operational challenges."
Claude Haiku 4.5
5 · The fix
One sentence per step recovers most of it.
We reran the chain two other ways. First, with one extra line in every step: "Keep every figure, caveat and condition from the text exactly as stated, and do not add anything that is not in it." Second, skipping the chain: writing the final 100-word client update straight from the original note.
6 · Try it yourself
Any document, any model, every step
Pick a document and a model, then step through the chain. Each step is a fresh AI call that sees only the previous step's text, as in a real pipeline. The checklist on the right shows what happened to each original fact.
Original facts
Intact: present and correctDistorted: changed, or only partly keptMissing
7 · Model by model
No model was safe. The gap between best and worst was wide.
Share of facts still intact after each handoff. Each row is nine chains (three documents, three runs). The last column is a cross-check that uses no AI judge: a plain text search for the original's key figures in the final message.
8 · Why this matters for media
Every newsroom using AI is running this chain
A wire story, press release or analyst note is rewritten into an article, a newsletter item, a push alert and a social post, more and more of it by AI. Each version reads well on its own. The damage only shows when you compare the last one with the first, and the first thing lost is the line that says how far to trust the story.
1. Lock the fragile linesCaveats, conditions and "not yet" statements should travel as fixed text that no step may rewrite. Numbers mostly look after themselves; the sentences that limit them do not.
2. Keep the source in reachLet every step see the original, not only the previous version. Writing from the original kept about twice as many facts as writing from the fifth copy.
3. Tell each step what must surviveOne instruction to preserve figures, caveats and conditions did more than switching to a bigger model. Then test it: run your own chain and compare the end with the start.
This page is that test. Give it a publisher's real content chain and it shows which facts break, at which step, and which model breaks them least.
How we did it
Method and limits
Method
Three invented documents (an analyst note, a drug trial summary, a customer survey), each with 9 to 11 checkable facts. The companies, drug and survey are fictional, so no model could know them.
Six handoffs: manager summary, client email, newsletter bullets, news paragraph, morning briefing, client update. Each step sees only the previous step's text.
Six models: GPT-4.1 mini, GPT-5.1, Claude Haiku 4.5, Claude Sonnet 4.5, Gemini 2.5 Flash, Gemini 2.5 Pro. One model runs a whole chain.
Scoring: after every step, GPT-4.1 compared the text with the original and marked each fact intact, distorted or missing. A plain text search for key figures was run as an independent check.
Limits
Small sample: three runs per document and model. The overall pattern is clear; the exact model ranking could shift with more runs.
The judge is itself an AI model from one of the tested vendors. "Distorted" includes facts that were only partly kept, such as a figure that lost its comparison.
Several steps ask for a shorter text, so some loss is expected. The direct rewrite to the same final length is the fair comparison, and it kept far more.
The chain is one plausible pipeline, not every pipeline. Different steps and prompts will lose different things.
Models were called through an enterprise gateway with default settings. Newer versions may behave differently.