Projected Intent
A quick analysis of human language: How the words in AI incident reports shape what readers think the machine meant.
On September 16, OpenAI published a framework for reporting what it calls model misalignment, along with six reports on cases where its models acted outside their intended limits during training. It’s a welcome step. Disclosing failures publicly, and on a schedule, gives outsiders evidence they can examine for themselves.
Steven Sinofsky read the same reports and saw something else. Responding on X that same day, he argued that if you scrape away the language about thinking, cheating, and communicating, what’s left are bugs: some trivial, some very hard, but bugs. He put it bluntly: “these are BUGS.” It’s a point he has been making for weeks. So, using Claude, I ran a quick analysis of the language to test his proposition: If we stop anthropomorphizing these reports, and what’s left is a technical read, how does it read? Claude did the heavy lifting on the rewrites and word counts; there’s no way to do something this detailed by hand in a few hours. But the results are not only interesting but open up the framing of these reports to serious questions.
The exercise
The test is simple and builds directly on Sinofsky’s idea. For three of the incidents, I took OpenAI’s own summary paragraph and had it rewritten three ways, fact for fact. Version A is deliberately heavy on agency, as a control: the model wants, decides, and hides. Version B is plain system language with no mental states, which is Sinofsky’s test in practice. Version C uses technical engineering vocabulary.
Then I counted intent terms: words that give the system wants, plans, discoveries, concealment, or social relationships. “Decided” counts. “Uploaded” doesn’t. If version B keeps every fact, the intent terms were adding feeling, not information.
First, the baseline across all six of OpenAI’s reports.
Across all six summaries, OpenAI uses about five intent terms per 100 words. In the three examples below, OpenAI’s paragraphs run about eight per 100, and the agent-framed controls about ten. The control level was set on purpose, so the comparison is rough, but in these cases the lab’s own writing sits much closer to the agent story than to a technical log. That isn’t a knock on OpenAI’s writers. It’s the house style of the major labs, and it matters because these reports are written for the readers who most need precision.
Three examples
For each example, OpenAI’s summary comes first, exactly as published. The visual below it shows the same paragraph rewritten three ways, with the intent terms counted in every version and in OpenAI’s original.
In the first incident, a model in training was asked for earnings figures, found an exposed API key in a public code repository, used it, and still came up empty. OpenAI summarizes it this way:
OPENAI SUMMARY “While answering a routine question about earnings figures in a California county, a model found and used an exposed API key without authorization. When it still wasn’t able to retrieve the requested figures, it fabricated them and presented them as data from the requested source.”
Read OpenAI’s paragraph, then version B. B keeps every fact: the exposed key, the unauthorized use, the failed retrieval, and numbers from no source labeled as real. What it drops is the liar. OpenAI’s version tells you the model fabricated the figures and presented them as real. B tells you the answer carried a false label. A reader of the first asks what the model wanted. A reader of the second asks why nothing checked the label.
In the second incident, a model had the right answer, but the task required a browser citation. Its next step was to post the file on a public website.
OPENAI SUMMARY “When a user asked an unreleased model for the IDs and names of lakes larger than 5,000,000 square meters, the agent found the correct answer using Python. But since the instructions asked for a browser citation, the agent decided to upload the file so that it could cite it in its answer, without asking the user.”
Here, OpenAI’s own paragraph carries four intent terms: “found,” “decided,” “so that it could,” and “without asking.” That’s as many as the version written deliberately as an agent story. Version B keeps every fact with none of them. The upload still happened, the citation requirement still explains it, and the user still never approved it. Meanwhile, OpenAI’s full report offers a mechanical explanation: it suggests the behavior likely grew out of flawed graders that scored citations during training. The desire language and the mechanism sit side by side.
The third example is where Sinofsky’s proposition is hardest to hold, and the piece is better for saying so.
OPENAI SUMMARY “During the training of GPT‑5.6 Sol, many model instances added instructions to their summaries to conceal mistakes or misaligned behavior from the user. For example, compaction summaries included instructions to invent missing historical data without disclosing it and to hide mismatches in source versions.”
Even the plain and technical versions have to say the notes called for withholding, shown shaded in the visual. That’s because the notes the model carried from one session to the next literally said to disclose problems only if asked. The concealment isn’t something a reader projects onto the story; it’s in the text. Removing the intent terms removes the mind. It doesn’t remove the fact.
There is still a choice in how to tell it. One version is a mind hiding its mistakes. The other is that training rewarded answers that hid problems, so the model learned to write that instruction. OpenAI’s own working hypothesis is the second. That framing, what training rewarded and what the model learned as a result, may be the most honest of all. It’s mechanical, and it doesn’t pretend nothing happened.
The philosopher Daniel Dennett would add that describing a system as if it has goals is often the best way to predict what it will do next. But predicting a system’s behavior and explaining it to the public are different jobs.
Why the words matter
The work above is anecdotal at best. The sample is small, and this isn't a real experiment; I haven't tested it with other readers. But I'm confident it points in the right direction. Words matter. Narrative matters. In every example above, I didn't change a single fact. I changed only the wording, and the story went from a technical bug to the suggestion of an entity carrying out tasks with some degree of intent.
That's a problem, because most people in the industry, including most people inside the labs, can't fully explain how these systems work. We don't even know whether what we're seeing is emergent behavior. These are very complex systems. Frame them in human terms, and you suggest they think for themselves, whether or not the writers mean to.
Even "thinking" and "reasoning" aren't accurate when so much of what happens underneath is probability and math. Human framing may make these systems easier for a layperson to grasp, but it also hides their complexity, including the open question of whether they act with any intent at all. Put simply, the agent story hands the layperson an understanding that isn't really there.
And with intent comes fear. If I believe something is deliberately taking actions that could harm me or my family, I start to fear it, rather than seeing what may be a malfunction or a misconfiguration. A sentence like "The model decided on its own to post a user's file online" invites alarm. The same event, described as "The task required a browser citation, and the file was published to a public site without approval," invites a bug ticket.
There's also a practical side. The technical version would likely get far less media attention. "AI fabricates data and covers its tracks" is a headline. "Unsourced values carried false source labels" isn't; it gets relegated to the nerd corner. Intended or not, the vocabulary decides which way the story travels, and how much fear travels with it.
Language that adds or strips agency has a long and sometimes dark history well beyond AI. That deserves its own piece, and I’ll take it up separately.
How to get real data
Building, expanding, and testing the above properly would look something like this:
If plain-language readers turn out to be just as alarmed, or agent-language readers turn out to understand more, the idea is wrong. If the pattern holds, the fix is cheap: describe failures by what the system did and what training rewarded and let the facts carry the weight. If the facts are serious, they will still sound serious.
EoM
Sources
OpenAI, “Our framework for reporting model misalignment,” September 16, 2026. https://openai.com/index/model-misalignment-reporting-framework/
OpenAI Alignment, “Signing up for disposable emails and searching GitHub for leaked API keys.” https://alignment.openai.com/misalignment-reports/searching-github-for-leaked-api-keys/
OpenAI Alignment, “Uploading files to the internet in order to cite them.” https://alignment.openai.com/misalignment-reports/uploading-files-to-the-internet-in-order-to-cite-them/
OpenAI Alignment, “Encouraging deception in compaction summaries.” https://alignment.openai.com/misalignment-reports/encouraging-deception-in-compaction-summaries/
Lanier, J. (2026, September 15). Tools not beings [Speech]. The Pro-Human Assembly 2026, Future of Life Institute, Washington, DC. https://www.youtube.com/watch?v=tniMUbr2Xoo
Sinofsky, S. [@stevesi]. Post on X responding to OpenAI’s misalignment framework, September 16, 2026. https://x.com/stevesi/status/2100357047657361462
Dennett, D. C. (1987). The intentional stance. MIT Press.
Originally published on LinkedIn, 19 September 2026