Ask the AI for the odds, and the essay goes away

Ask a chatbot whether a claim needs a site visit and you get four paragraphs, when what you needed was a yes or a no and how sure it is.

You paste in a first notice of loss and ask a simple question: does this one need a surveyor on site? The chatbot writes four careful paragraphs and leaves the decision where it started. It did what it was built to do. A generative model produces text, and it keeps going until the answer sounds complete.

Liquid AI has built a different kind of model for questions with a fixed set of answers. Its d1 takes text, images or both, plus a question, and returns a probability for each possible answer without writing a word of text, the company says (Liquid AI). It handles three kinds of question: yes or no, a choice among labels, and a score on a scale. A text decision takes 200 to 300 milliseconds. Liquid AI puts the cost at 19 to 200 times below the two leading chat models it tested against.

The image update is the part to watch. Sorting good parts from defective ones on a production line, d1 reached 85% to 97% accuracy. "The most interesting part is that the model was never trained for this," the company writes. You do not need d1 to borrow the idea. Most chatbots will answer as a decision if you give them the allowed answers and ask for a probability on each one, though the numbers they give still need checking.

On a claims desk, most first questions are decisions: site visit or desk review, repair or total loss, refer to the fraud team or pay. A probability gives you a threshold to manage: pass files above 0.8, review the middle band by hand, and question anything the model calls certain. A threshold only means something once you have tested it on files where you already know the answer, which is the job of the prompt below.

Run it: score 20 closed files, then check the scores

Use this with any chatbot. Pick 20 closed claims where you know the outcome, remove names, plates and policy numbers, and give each file a short ID.

You are a claims triage assistant. Answer as a decision, not
as an essay.

The question for every file:
[e.g. Did this claim need a surveyor on site?]

The only allowed answers:
[e.g. YES / NO]

For each file I give you, return one line in this format:
File ID | Answer | Probability of that answer (0.00 to 1.00) |
The single fact in the file that moved the probability most

Rules:
 1. Use only what is written in the file. Do not assume facts.
 2. If the file does not hold enough to decide, answer
    INSUFFICIENT and name the missing fact.
 3. Do not explain beyond the one fact.

Here are the files:
[paste the 20 closed files, details removed, each with its ID]

When you have scored all 20, stop and wait. I will then give
you the real outcome for each file. Build a table comparing
your answers with the outcomes, grouped into three probability
bands (below 0.60, 0.60 to 0.80, above 0.80). Tell me in which
band your answers were wrong most often, and which facts you
relied on in the files you got wrong.

Prompt written for this edition. It has not been published elsewhere. The decision-model concept is from Liquid AI's d1 announcement of 5 October 2026.

Claims job ads still describe the job before AI arrived

Insurers are putting AI into claims, but the ads for the people who handle claims still describe the job as it was before.

Pull up the last claims-handler vacancy your company posted. It probably asks for policy knowledge, customer skills and good judgment. It probably says nothing about the AI tools that triage or score files in more and more claims teams, or about what the handler should do when a tool gets it wrong.

Shift Technology, which sells AI to insurers, checked how common that is (Shift Technology). It read 100 job ads on the career pages of US insurers. Of 67 ads for claims-handling and claims-operations jobs, three mentioned AI. Six ads for underwriters and six for fraud and recovery investigators did not mention it at all. The ads that did talk about AI were almost all for the people who build it.

The ads still ask for people: 76% of the operational ones want human judgment. What is missing is the work that sits next to an AI tool. One ad in ten mentions handling exceptions, escalating a file or checking quality. Shift's conclusion: "The next phase of AI readiness is human." The sample is small and US-only, and Shift says it is not a full picture of insurance hiring.

The ad matters because it sets what a person is hired for, trained in and judged against. If the ad leaves out the AI, nobody is accountable for checking it. For a claims handler, three lines are missing: check the system's first decision before it reaches the customer, escalate a file the system got wrong, and explain an AI-assisted decision to the customer in plain words. If those lines are not in your own job description, ask for them at your next review. If you hire, put them in the next ad.

ChatGPT drew the cartoon, then signed a real artist's name

If a chatbot adds a real person's signature without being asked, a signed image sent in by email or app is weaker proof than it looks.

In May, a stranger messaged the cartoonist Brendan Loper on Instagram: his pen name was on an image ChatGPT had made. Loper then found more examples in a Reddit forum, where users posted New Yorker-style cartoons the chatbot had produced. A viral cartoon of Dolly Parton, made when a fan asked for a New Yorker-style cartoon, carried the signature "BLOPER."

Andrew Deck at Nieman Lab documented more than 15 New Yorker cartoonists whose signatures appeared on ChatGPT output, used without permission or payment (Nieman Lab). "It felt very much like a violation of my personhood," Loper said. "When they hacked my credit card, they didn't dress up like me," said Joe Dator.

OpenAI thanked users for flagging bugs and unintended behavior, and declined to say how its models learned the signatures. After Nieman Lab contacted it, ChatGPT began warning that such prompts may breach its guardrails on similarity to third-party content. At publication, some generic cartoons still came out signed. Condé Nast said it has never permitted any AI developer to train on its cartoons.

Two reads for an insurer. The first is evidence. A claim file is full of signed and stamped documents: repair invoices, surveyor reports, police letters. The model added the signature because signatures were part of the pattern it learned, and nothing in that mechanism is specific to cartoons. Confirm key documents with the issuer, through a channel the claimant does not control.

The second is liability. A business that publishes AI images carrying a real person's name has a content exposure, and its media liability wording should say whether that is covered.

The rule about being cruel to Claude arrived with a rule about your claims decisions

The headlines were about rudeness, and the part you should read covers who decides a claim and who you may watch.

Most people who use a chatbot at work have typed something sharp at it after the third wrong answer. That is not what Anthropic's new rule targets. From 12 November, its usage policy prohibits "sustained and needless abusive or cruel behavior toward our models," in extreme cases where a user acts cruelly again and again with no apparent purpose (Anthropic). Ordinary frustration and pushback are out of scope. The Register notes there is no established evidence that Claude experiences distress (The Register).

The same update, published on 8 October, changes rules much closer to an insurer's work. Tracking people without their consent is prohibited, whether in real time or by analyzing data already collected. Consented tracking, such as fraud monitoring, stays permitted, and Claude may not be used to build or improve surveillance tools.

Where Claude drives physical equipment that could injure someone, "a qualified operator must be able to observe the equipment and stop it if needed." Where Claude affects health, legal rights, finances, livelihood or access to essential services, a qualified human must be able to review and change its recommendations, and the affected person must be told AI was used. Anthropic says these requirements are unchanged and that the update specifies which recommendations they cover.

A claim decision affects someone's finances, so a Claude-based tool on a claims desk sits inside that last rule. Before 12 November, check three things. A named person can change every recommendation the tool makes. The customer learns that AI was involved. And any investigation work, such as checks on a claimant's public activity, has the claimant's consent or falls in a use Anthropic's policy allows, such as fraud monitoring, as well as meeting your own legal and privacy rules. If your fraud team uses a general chatbot, that third check is theirs to run this month.

A report that animates itself, an interviewer that waits for your answer, and a ring that hands the job off

One explains your work to others, one rehearses you for a hard conversation, and one takes a request without a screen.

Your people overrule the agents, your brokers have no AI policy, and a worm runs on TV boxes

Each story below asks a different desk the same question: who checks the machine, and what happens when nobody does?

The agents your staff keep overruling

"Cutting costs is the least imaginative use for AI," says Franklin Manchester, global strategic insurance advisor at SAS. SAS surveyed 2,699 decision-makers in 28 countries in April 2026 for its second Data and AI Impact Report, with research insights from IDC (SAS). Among insurers, 76% expect to increase AI investment next year and 84% expect a positive return. Across all industries, nearly 80% override more than 10% of AI agent decisions, and nearly a third override more than 30%. The top reason, at 34.5%, is a lack of explanation.

Among insurers the report rates as leaders, 59% expect strong or high returns, against none of the laggards. If your teams overrule an AI agent often, record the reason every time. Over a quarter, that record shows which overrides need a better explanation from the model and which need a different model.

Half the agencies selling your policies have no AI rules

Picture an account manager at a small agency pasting a client's renewal schedule into a free chatbot. Nobody told her not to. 56% of US independent agencies have no written AI policy or guidance, according to the 2026 Technology Trends Report from the Big "I" Agents Council for Technology (Big "I" ACT). Mike Allton argues that carriers should set the rules before an errors and omissions claim forces the issue (Insurance Thought Leadership).

He asks carriers for four things: a written acceptable-use policy as a condition of appointment, a yearly sign-off alongside proof of E&O cover, a named owner for every tool, and a template policy agencies can adopt. If you appoint brokers or agents, the first and the last cost little to add to your next appointment pack.

A worm that runs on $55 TV boxes

Anthropic's Frontier Red Team ran a small AI model on 15 off-brand Android TV boxes, about $55 each, then tested whether models could copy themselves across a sealed test network (Anthropic). In a simplified test with the hacking steps kept trivial, one open-weight model, GLM-5.2, spread to all eight nodes in four of five trials. If small models keep improving at their historical rate, the team expects them to reach that level by the end of 2026.

"A replicator like this would be challenging to shut down, since there's no centralized point to shut off," the authors write. For a cyber or home underwriter, this is an accumulation question. One replicator spreading across cheap home devices would reach many policyholders at once, through devices nobody lists on a proposal form. Check how your cyber and home wordings treat a botnet event, and whether connected-device security appears anywhere in your underwriting questions.