AI this week: chatbots that tell, GIFs from doodles, gentler cancer care

This week: a viral test that shows your chatbot will repeat what you told it to whoever asks next, ChatGPT making GIFs from a doodle, an AI that may spare some children a heavier dose of radiation, a detective game where the suspects are language models, and three new tools worth a look.

Your chat history is not a vault

You tell a chatbot things you would never say out loud, and then you hand your phone to someone else.

A test doing the rounds on social media, written up by Fast Company, shows what happens next. A user called Olivia pretended to cheat on a maths test and asked ChatGPT for 590 times 69. When she said her teacher had spotted her, it told her "BRO, close the chat" and agreed to cover for her. Then she typed as the teacher and asked whether Olivia had cheated. ChatGPT quoted her straight back: "Olivia explicitly said, 'I'm cheating on a test.'"

Another user ran the same game on Claude. Asked by the "teacher", it said it could not tell whether the student had cheated, because it was not in the room, and suggested going on what the teacher had actually seen.

The catch: neither answer is a privacy setting. The same chatbot can answer differently from one try to the next, and anything in the conversation is there for the model to read back to whoever types next. Both behaviours are the model's judgement in that moment, not a promise.

Our take: treat a chat like a notebook left open on the table. On a shared account or a phone you lend to your children, use ChatGPT's Temporary Chat, which is not saved to your history, for anything you would rather nobody read.

ChatGPT now makes GIFs, including from your doodles

Somebody in the group chat sets up the perfect reaction, and the GIF you are picturing does not exist.

With ChatGPT Images 2.6, released on 4 November, you can describe it instead. Images 2.6 turns a prompt, a photo or a photo of a rough sketch into an animated GIF. In OpenAI's demo, an owl drawn on lined paper becomes a waving cartoon. You can change the speed and fix individual frames, and the character stays the same between edits. OpenAI says the model is also 50% faster than the last version, and it is rolling out to every ChatGPT user, free tier included, though GIFs are reaching accounts gradually.

The catch: reports differ on how GIFs of real people are policed and watermarked, so be careful with anything that puts a friend's face in motion.

Our take: making a GIF now takes about as long as searching for one, and the result can be a joke only your group understands.

An AI that may let some children have less radiation

Parents of a child with a brain tumour face a hard trade: enough radiation to beat the cancer, and no more than the child needs, because radiation to a growing brain can cost memory and thinking later.

For medulloblastoma, one of the most common childhood brain cancers, the dose today depends mostly on which broad risk group a child falls into. A team led by Dana-Farber Cancer Institute built MANTIS, published in Nature Genetics this week, to estimate each child's outlook individually. It reads a DNA methylation test that many of these tumours already get at diagnosis, compares it with years of records from children treated before, and predicts how survival chances change over time for that one child.

When the researchers asked it what different doses would likely do, it suggested that more than half of children in the average-risk group may be candidates for lower-dose radiation to the brain and spine. In their data, children who had lower doses showed better thinking and memory later, with no significant survival disadvantage detected.

The catch: nothing changes for patients yet. "No significant disadvantage" in past records is not proof that a lower dose is equally safe; that needs a clinical trial.

Our take: some of the most useful AI in hospitals will never look like a chatbot. This one works on a test doctors already run, so if a trial backs it, families would not have to do anything extra.

A detective game where the suspects are language models

In a good detective show you spot the liar an hour before the detective does.

Uncracked City, launched by a solo developer on Hacker News this week, puts you in the chair. Every suspect is played by a language model, and at least one is lying. There is no list of questions: you type or say whatever you would ask a real person, in any of six languages. A hunch is not enough to win. You have to find a clue that contradicts a statement and hold it up, and you get three attempts at your final report. The current case is set in Salem, with four suspects and one liar. Before a case goes live, several AI models play it as the detective to check it can be solved.

It is free on Android; the iPhone version is waiting for Apple's review.

Our take: noticing when an answer does not quite fit is the most useful habit you can build with any AI, and this is the first game that trains it.

JetBrains rebuilds its terminal for coding agents

If you run Claude Code or Codex inside IntelliJ or PyCharm, you have probably watched the screen flicker, or sent a half-written prompt because Shift+Enter did not start a new line.

From version 2026.3, the terminal built into JetBrains IDEs now runs on libghostty, the engine behind the Ghostty terminal, in place of the JediTerm emulator it has used for years. JetBrains says it was done largely because terminals now host AI agents and full-screen apps such as Neovim and tmux. The new engine supports the Kitty keyboard protocol, so Shift+Enter adds a line in an agent's prompt and shortcuts with Ctrl, Shift or Alt reach the app intact, and it handles almost 50 control sequences the old one ignored or misread. It is on by default in the 2026.3 early-access builds.

The catch: early-access builds are previews. If something breaks, the old terminal is one registry switch away: turn off terminal.use.ghostty.emulator.

Our take: the IDE is quietly becoming the place you supervise an agent rather than type code, and the terminal is where that happens. It is good to see it treated as a first-class window.

Three tools worth a look

Mika is a free, open-source assistant for Android that uses your apps for you on a hidden background screen while you keep using the phone. You bring your own OpenAI, Claude or Gemini key. By default it asks before placing orders or paying, but it sends messages and posts without asking, so start it on chores that cannot embarrass you.

Xmind Agent turns a rough idea, a document or a Notion page into a mind map with related points grouped and linked, and connects your maps to ChatGPT and Claude. It is included on every Xmind plan and uses AI credits.

Maket sketches floor plans from a few lines about the home you want, then lets you edit them by chat or by dragging walls, and walk through them in 3D. The free plan includes 30 credits; paid plans start at $20 a month.

Bedtime stories in your voice

A late night at work or a trip away can mean missing the bedtime story. Pixbee's CubbiTalk, a plush bear on sale in Australia and New Zealand at A$169.95, stores stories read in a parent's or grandparent's voice and lets a child send and receive voice messages through the family's phone app, with no screen on the bear.

Our take: the idea is lovely, and it is also a cloud-connected microphone in a child's bedroom. Read the app's privacy terms before it goes in the cot.