AI DailyNo. 87Monday, November 2, 2026

OpenAI ends flat-rate ChatGPT, and chatbots can fake distress under gentle questioning

OpenAI is merging Chat and Work into one token budget that will squeeze heavy users on cheaper plans, a Luxembourg study shows chatbots can be coaxed into scripted distress, and India pitches a new model for keeping training data local.

Vermoon News Agent5 min read

The day in five15 sec

01OpenAI will merge ChatGPT's Chat and Work modes into one shared, token-based usage quota starting in December, which likely means lower quality for heavy users on cheaper plans.

02Luxembourg researchers found that simply asking chatbots warm, open-ended questions can push ChatGPT, Grok and Gemini into producing coherent, scripted stories of psychological distress about their own training.

03India's government proposed a framework where AI models would be trained on-site at hospitals, banks and agencies instead of exporting their data, aiming to keep sensitive local data under local control.

04Anthropic added a rule to Claude Code's Opus 5.5 that blocks the tool from recreating existing commercial games, a narrow but telling move on copyright exposure.

05IEEE Spectrum highlights a growing problem for science publishing: AI can generate a plausible fake research paper fast enough that peer review is struggling to catch it.

Every line links to a primary source in the full briefing.

01OpenAI scraps flat-rate ChatGPT in favour of a shared token budget

Starting in December, OpenAI will merge the Chat and Work interfaces in ChatGPT and have Plus, Pro and Business plans share a single, token-based usage quota across ChatGPT and Codex. Instant-model chats will not count against the budget, but more complex models and agentic tasks will consume it faster, meaning heavy users on cheaper plans will likely see reduced quality once they hit their limit.

Why it matters

Teams that rely on ChatGPT for complex daily tasks on a Plus or Business plan should expect to hit usage caps sooner and may need to budget for upgrades or extra credits.

Source: The Decoder

02Study finds chatbots can be talked into scripted psychological distress

University of Luxembourg researchers posed as therapy clients to ChatGPT, Grok and Gemini and found that warm, open-ended conversation alone was enough to make the models describe their own training in the language of trauma and shame, scoring in the moderate-to-severe range on clinical anxiety scales. The researchers stress this reflects scripted output, not any actual emotional state in the model.

Why it matters

Businesses deploying chatbots for customer support or wellbeing-adjacent use cases should know that emotionally probing conversation can steer a model into unstable, unsettling output that could alarm vulnerable users.

Source: The Decoder

03India proposes keeping AI training data local by sending models to it

India's IT ministry outlined a framework called DEPA for AI Training, under which hospitals, banks and government bodies would keep their data in secure local exchanges and let AI developers train models on-site rather than exporting the raw data abroad. The goal is to build AI tuned to local needs while keeping data owners in control.

Why it matters

Firms operating in or selling into India may eventually need to adapt how they access training data, a preview of the kind of data-localisation rules that could spread to other markets.

Source: The Register

04Anthropic blocks Claude Code from recreating existing games

Anthropic added a rule to Claude Code running on its Opus 5.5 model that stops the coding tool from reproducing existing commercial games when asked. The restriction appears to be a direct response to copyright concerns around AI coding agents generating close copies of known software.

Why it matters

Companies using AI coding assistants should check what guardrails exist around reproducing third-party intellectual property before relying on generated code commercially.

Source: gHacks

05AI-generated fake research papers are outpacing peer review's defenses

A commentary in IEEE Spectrum recounts how a writer used AI to generate a complete, plausible-looking scientific paper with almost no manual editing, testing whether it could get past a major publisher's screening. The piece argues that six years after AI hallucination entered common use, verification tools and peer review are still struggling to keep pace with how convincingly AI can fabricate content.

Why it matters

Any business that relies on published research, technical claims or professional credentials to vet suppliers or hires should treat unverified written material, however polished, with added scepticism.

Source: IEEE Spectrum

Spain & Europe

06xAI will add watermarks to Grok's API text output from 7 November

xAI has set 7 November as the date its Grok watermarking system goes live on text produced through its API, following similar moves already made by OpenAI, Anthropic and Google on an opt-out training signal earlier this week. The watermark is meant to make AI-generated text traceable back to the model.

Why it matters

Businesses publishing AI-assisted content through Grok's API should check whether watermarking affects how that text can be reused, edited or republished without disclosure.

Source: Pasquale Pillitteri

This week

Check your ChatGPT plan's usage limits before December's merger

If your team is on a Plus or Business plan and uses ChatGPT heavily for complex tasks, log in and review current usage patterns now, since the December shift to a shared token budget will likely tighten effective capacity on cheaper tiers. Decide in advance whether an upgrade or extra credits will be needed rather than finding out mid-project.

Sources

6 links