Imagine being held accountable for a promise you never made.
Six months ago, you discussed a possible delivery date in a meeting. Today, someone opens the meeting notes and asks why you missed the deadline.
The notes look clear. Your name is there. So is the date.
Except you never agreed to it.
This is one of my growing concerns about AI in the workplace: what happens when an AI-generated version of a conversation becomes the version everyone trusts?
I call this organizational bias: a company starts treating unchecked AI output as shared truth.
We already have a name for part of this behaviour. Automation bias is the well-documented tendency to trust what a system tells us without checking enough, and to discount information that contradicts it. My concern is what happens when that habit stops being an individual mistake and becomes how an organization remembers.
AI output starts shaping what people believe was said, decided, and promised.
A conversation becomes a commitment
Let’s take a simple example.
In a meeting, a project lead explains that September might be possible, provided testing goes well and two extra engineers are available.
The AI summary records a September delivery date.
The conditions disappear. The uncertainty disappears. A possible outcome becomes an apparent commitment.
A transcript and a summary fail in different ways. A transcript can capture the wrong words or assign them to the wrong person. A summary can capture the words correctly but lose the conditions that gave them meaning.
This is not a hypothetical weakness. In 2025, researchers at Utrecht University and Western University compared 4,900 AI-generated summaries of scientific papers with the originals. Up to 73% of the summaries from the most popular models turned careful, conditional findings into broader claims than the source supported. The models were nearly five times more likely to overgeneralize than human summarizers. And when the researchers explicitly prompted the models to be accurate, the summaries got worse, not better.
Scientific papers and meeting notes are different documents. The failure mode is the same: hedges, scope, and conditions are exactly what a summarizer is built to strip out.
Once that September date enters a project plan, it can appear in a management update and then a quarterly review.
Every repetition makes it look more established. Nobody necessarily checks where it came from.
The record is now evidence
Until recently, this was an internal problem. It is becoming a legal one.
On 28 August 2026, the Delaware Court of Chancery decided a boardroom dispute in which two written records of the same meetings sat side by side as exhibits. The official minutes said a defensive measure was adopted “to protect stockholders”. The AI-generated transcript captured the chairman saying it was “necessary in order for the board to remain in its position”. The court read both. The board lost.
Neither record was the whole truth. The minutes were curated. The transcript was raw. Together, they told a story nobody in the room had signed off on.
Meanwhile, AI notetakers themselves are in court. Class actions against Otter.ai over recording participants without consent were allowed to move into discovery in August 2026, and a similar case against Granola was filed in July. The practical lesson for every company, regardless of jurisdiction, is simple: an AI transcript or summary is a business record. It can be requested, quoted, and held against you, whether or not anyone reviewed it.
We can generate more than we can review
I use AI daily and see its value. It helps us write, organize information, and move faster.
But our ability to produce content is growing faster than our capacity to check it.
Meeting notes. Strategy documents. Requirements. Decision matrices. Another summary of another summary.
Each document takes seconds to generate. Understanding whether it accurately reflects the situation takes attention and context.
When people are busy, a polished document is easy to accept. Especially when everyone assumes somebody else has already reviewed it.
Stanford’s Social Media Lab and BetterUp Labs put a name and a price on this in 2025. They call it workslop: AI-generated work that looks like good work but lacks the substance to move a task forward. In their survey of US desk workers, 40% had received workslop in the previous month. Each instance cost the recipient close to two hours to untangle. They estimate the bill at roughly $186 per employee per month.
The most telling number is not the cost. It is that 42% of recipients said they trusted the sender less afterwards.
The cost reaches beyond accuracy
Imagine a leader using AI to turn a few ideas into a strategy document.
The document sounds confident. But it introduces priorities the leader never intended and dependencies nobody discussed.
The team now has to work out what the leader actually means.
Or imagine an AI-generated decision matrix recommending a supplier. The scores look precise, but the criteria and weightings were never agreed upon.
People start questioning the judgement of whoever shared the document. Colleagues spend time untangling the message. Trust weakens.
The person sending it saved time. Everyone receiving it inherited the checking.
“The person sending it saved time. Everyone receiving it inherited the checking. That is not productivity. That is a cost transfer with a delay.”
Why this is an organizational problem, not a tooling problem
It is tempting to treat this as a vendor issue. Pick a better notetaker, wait for the next model, and the problem goes away.
The evidence says otherwise. When the European Broadcasting Union and the BBC tested more than 3,000 answers from ChatGPT, Copilot, Gemini, and Perplexity across 14 languages and 18 countries, 45% contained at least one significant issue: wrong or missing sourcing, factual errors, or lost context. The failures were consistent across tools, languages, and territories. They are systemic.
Tools will improve. But the structural problem remains: an organization that has no habit of checking will absorb whatever error rate its tools have, at whatever scale it generates documents. The fix is not a better model. The fix is a review habit that survives the next model.
What I would change
We need a few practical habits:
- Confirm commitments. Ask the people involved to verify decisions, owners, deadlines, and conditions before the notes are circulated. If a date has a condition attached, the condition is part of the date.
- Keep proposals separate from agreements. Something discussed in a meeting is not automatically approved. Notes should say “discussed” when it was discussed and “agreed” only when it was agreed.
- Check what drives a decision. In a decision matrix, review the evidence, criteria, and weightings, not just the ranking at the bottom.
- Make corrections visible. If incorrect notes have reached a project plan, fix the plan too. A correction that stays in the source document while the copy keeps circulating is not a correction.
- Own what you share. Read it and make sure you can explain it in your own words. If you cannot, it is not ready to send.
A second AI model can help spot problems, but agreement between two models is no guarantee of accuracy. Important claims still need checking against the original evidence and the people involved.
For me, this is a leadership responsibility. Teams need time to review important output and room to challenge it, even when it comes from senior management.
Before an AI-generated document becomes company truth, someone needs to answer a simple question:
Does this reflect what we actually meant and agreed?
Frequently Asked Questions
Isn’t this just automation bias with a new name?
Automation bias describes an individual over-trusting a system. Organizational bias is what happens when that behaviour is repeated across a company and the unchecked output becomes the shared record. The individual mistake is recoverable. The organizational version compounds with every copy.
Should we stop using AI notetakers?
No. Use them, and treat the output as a draft rather than a record. Decide who reviews it, how quickly, and what gets confirmed with the people involved before it is filed or forwarded.
Who should own the review?
The person who shares the document. Sending an AI-generated summary is an act of authorship. If you send it, you are vouching for it.
Does the AI transcript replace the minutes?
It should not. Minutes are the agreed record. A transcript is raw material. The Delaware case shows what happens when a company keeps both without deciding which one is authoritative: a court decides for you.
AI has made it cheap to produce a confident version of events. It has not made it cheap to produce a true one. Until the checking catches up with the generating, the most important sentence in any AI-generated document is the one someone adds by hand: reviewed and confirmed by the people involved.
If you are working through how to introduce AI into your organization without losing control of the record, get in touch.