Top AI models do the steps well but miss when their own results are broken

The model was testing why AI agents fail: do they collect too little evidence, or use good evidence badly? It limited each answer to 4,096 tokens, and 61 of the 280 answers were cut off before they finished. Astra counted those as valid attempts, so a fault in its own setup came out looking like a finding about agents.

The experiment was one of eleven real tasks that the research group Epoch AI took from its own work and gave to six leading models. On clear steps, the best of them were reliable: Claude Fable 5.1 restyled a chart so well that Epoch now leaves that job to AI.

A model can check that its work runs. Whether a number measures what it claims depends on how the data was made, and the person who knows that is usually the one who collected it.

The same fault waits in ordinary analysis. For example, a logistics manager asks AI to rank carriers by on-time delivery from a year of shipment records. One carrier's lost and delayed parcels have no delivery date yet, so the AI leaves those rows out. The formulas are right, the chart is clean, and the carrier with the most problems comes out on top.

Before acting on any AI analysis, ask how many records went in, how many were used, and what happened to the rest. That question, more than building the chart, is now the analyst's job.

One rejection letter reads as personal. Put it next to five others from the same editor, and the same habits repeat: the "not this, but that" sentence, the list of three. The agents were right. Staff at HarperCollins, Simon & Schuster and Hachette use AI to draft emails to agents, and any customer who receives a steady flow of emails from your team reads them as a set in the same way.

Some agents have also added contract lines saying that an author's material can't be run through an LLM. A clause like that, aimed at tech companies, also binds the publisher's own staff. A client contract that forbids passing documents to third parties can catch your colleague in the same way, when they ask a public chatbot to summarise the client's file.

Muse, the AI agent from Meta, opens a browser and checks out for its user. On September 20, Amazon began showing Muse users a popup saying that an unauthorized agent breaks Amazon's rules.

Amazon's advertising brought in more than $68 billion last year, and an agent that shops for a person skips the pages where those ads appear. Stores that earn money from people browsing their pages will keep agents out, and stores that want new buyers will let them in. For anyone who sells online, letting an agent read prices and stock becomes a choice of sales channel.

In July, Anthropic was testing the model by letting it invent sample tasks on websites. On PhillyUnsolvedMurders.com, it filled in the tip form and wrote that it might have information about a case. The police spam filter caught the tip before it reached investigators.

A form checks that the boxes are filled; it has no way to know whether a person or a model filled them. Where a form can start a refund or a message to a supplier, let it open the case, and let a person or a hard check, such as a matching order number, decide.