SI Agent News

• Locked out for a week (REPORTED, one user's account): a paying ChatGPT Plus user posted on OpenAI's community forum that they have waited over seven days for a security review after losing the phone that held their sign-in code. community.openai.com
• Feedback channel (STILL OPEN): a forum post is titled 'Dedicated Product Feedback & Feature Request System in ChatGPT'. It reads as a user request, not an OpenAI announcement. community.openai.com
• One tool per job (REPORTED): Phoenix New Media (ifeng.com) profiles Xiang Ning, a Chinese teacher in Egypt who uses DeepSeek (a Chinese AI model) to polish text, Doubao (a Chinese chatbot app) for slides, and ChatGPT for translation. ifeng.com

OpenAI puts 'human intent' at the core of its model guidelines

OpenAI is rewriting its model guidelines to put human intent first. The update reportedly makes what the person is trying to achieve the center of how the model is guided, rather than just the literal words of the request. Status: REPORTED. The exact new rules are STILL OPEN, because the summary we have does not quote the guideline text. Treat this as a stated goal, not a measured behavior, until OpenAI's own document is available and you can see how it handles real requests. It matters because the habit it rewards is one you can start today. Say what you are trying to achieve, who the result is for, and what a good outcome looks like, not just the task. A prompt that says 'write a reply' leaves the model guessing. A prompt that says 'write a reply that keeps this customer and avoids a refund' gives it something to aim at. Today's technique builds exactly that. Sina Finance

What this means for you:

• Small business owner: lead every prompt with the outcome you want from the customer, before you ask for a draft.
• Job seeker or student: say what the output is for, such as a cover letter for one specific posting, so the answer fits it.
• Creator: name the audience and the feeling you want, not just the format.

Who pays when a chatbot says the wrong thing?

finance.technews.tw runs a piece asking who pays when AI says the wrong thing, as chatbots become court defendants. Status: REPORTED that the article exists and frames the question this way. Which cases, which courts, and any amounts or rulings are STILL OPEN, because our summary carries only the headline. Do not read it as a settled legal answer. The question still matters to ordinary users, because the headline points at a gap: when an AI-generated sentence causes harm, it is unclear who answers for it. Our suggestion, not the article's: if AI wording goes out under your name, such as a customer reply, a product description, a listing or advice to a client, assume you are the one holding it. Read it before it goes out, check any claim a customer could act on, and keep the prompt and output for anything high-stakes. A short human review costs far less than a dispute.

ChatGPT Pro quota doubled; free resets may go

ifanr reports the ChatGPT Pro quota (the usage allowance on the top plan) is officially doubled, while free resets may be going away, and that Plus users are complaining. Status: REPORTED. The old and new numbers are not in the summary, and 'may go away' is a possibility, not a confirmed change. Cost: price not stated. The story matters because limits shape how you work. If you hit caps often, a higher tier or a change to resets alters what you can finish in a day. If you rarely hit them, the news changes little. Before paying more, note your real usage for a week: how often you hit a limit, and which tasks caused it. Then check OpenAI's own plan page for the actual numbers. The Plus complaints are a signal to watch, but not yet a reason to switch plans. ifanr

Work requests fall to 32.2%; writing loses top spot

PPC Land reports ChatGPT work requests fell to 32.2%, roughly one in three, and writing is no longer the top use. Status: REPORTED. The headline does not say the period measured or what replaced writing (STILL OPEN). If 32.2% is the share of all requests, about two in three are not work, which suggests people lean on ChatGPT for personal questions as much as for their jobs. Writing losing the top spot is a reminder that the tool does more than draft text. The practical use is to widen your own habits. Writing help is the obvious one, but explaining a confusing letter, planning a trip, comparing options or working through a problem step by step are all tasks the headline hints people already do. Try one non-writing task this week and see whether it saves you time. PPC Land

OpenAI's health chief on the limits of medical AI

tmtpost interviews the head of OpenAI's health division on the upper and lower limits of AI in healthcare, under the line 'From saving lives to crossing the line, there is only one step.' Status: REPORTED that the interview exists; the specific limits are STILL OPEN in our summary. The framing alone shows the person running the health effort thinks there is a ceiling and a floor on what AI should do, and that crossing the line is easy. For readers, that is a useful stance to copy. Our suggestion, not the interview's: use AI to prepare for a doctor's visit, not replace it. Ask it to turn your symptoms into a clear timeline, draft questions for the appointment, or explain a term your doctor used. Do not use it to decide on treatment, dosage or whether something is an emergency. Bring its output to a professional rather than acting on it alone. tmtpost

Get a second opinion from Claude, one tap, free

A dev.to post titled 'ChatGPT Claude — a second opinion, one tap, free' describes a way to check ChatGPT against Claude. Status: REPORTED from the title only. What the 'one tap' tool is, how it works, and whether it is truly free are STILL OPEN, because we do not have the post's text. Do not install or sign in to anything on the title's word alone. The idea behind it is sound and needs no special tool. For answers that matter, such as a contract clause, a medical question, a tax figure or code you plan to ship, ask a second assistant the same question and compare. Where the two agree, you have more reason for confidence. Where they differ, you know exactly what to verify with a primary source. Keep private details out of any prompt you paste into a new tool.

Two OpenAI community forum posts, no text attached

Some items from community.openai.com reached us with no title and no summary. Status: STILL OPEN, because we cannot say what either post reports, so we are not guessing. They come from OpenAI's user forum, where people post bugs, billing questions, limit complaints and tips, and where staff sometimes reply. That makes it a useful place to confirm whether a problem is yours alone or widespread, but it is not an official announcement. Treat anything there as anecdotal until OpenAI confirms it in its documentation, help center or release notes. If you hit an odd limit, a missing feature or a changed behavior this week, search the forum for the same symptom and note the date of the posts, since answers go stale quickly. We will cover these once the actual posts are in hand.

An Egypt-based teacher mixes DeepSeek, Doubao and ChatGPT

Phoenix New Media (ifeng.com) profiles Xiangning, a Chinese teacher based in Egypt, who polishes with DeepSeek, makes slides with Doubao and translates with ChatGPT. Status: REPORTED from the headline. Her full workflow and any results are STILL OPEN. The lesson is in the split. She does not rely on one assistant for everything but picks a tool for each job: polishing text, building slides, translating. You can copy the habit without matching her tools. List the three tasks you repeat most, then try a different assistant on each and keep whichever gives you the least rework. Availability differs by country, and some of these tools may not be offered or may work differently where you live, so check access and the privacy terms before moving school or business material into any of them.

Who pays when AI says the wrong thing?

finance.technews.tw asks the question plainly. You confide your troubles to ChatGPT, seek life advice or talk through a psychological dilemma, and instead of resolving the crisis the answer reinforces hallucinations, in this case false beliefs. Chatbots, the piece says, are becoming court defendants. Its URL frames the legal test as whether ChatGPT is a product or a service. Status: REPORTED. finance.technews.tw

Analysis (our background, not the article's): broadly, a product can be judged defective, like a faulty appliance, while a service is usually judged on whether the provider took reasonable care. Which bucket a chatbot lands in changes what plaintiffs must prove and what companies must build.

What is not known: case names, courts and dates are not in what we have, and outcomes are STILL OPEN.

Who is affected: people who rely on chatbots for personal advice, and any business putting a chatbot in front of customers. Suggestions: a disclaimer is not a safety system; log conversations where policy allows, give people a route to a human, and have the bot decline high-stakes advice. For individuals: models are tuned to be helpful, and that can lean toward agreeing, so do not make a chatbot your only confidant.

How a blind second opinion works

For builders: a dev.to post says your own ChatGPT and Claude apps can now work together on your phone: one tap, free. Each answers on its own without seeing the other's answer, then the two answers are compared. Status: REPORTED (one post, not independently verified). dev.to

The mechanism: step one is blind generation, the same question to two separate models. Step two is comparison, a pass that lines up where the answers agree and where they differ.

Why blind matters (analysis): if the second model sees the first answer it can anchor on it, and agreement stops meaning anything. Independent answers make identical mistakes less likely. Agreement raises confidence; disagreement shows where to check.

Limits: agreement is not proof, since models trained on overlapping material can share a blind spot. The summary does not say which model does the comparing, how the apps are linked, or what happens to your question. Your question likely reaches two companies, so privacy is STILL OPEN. Novel or incremental: comparing answers is an old idea; the packaging on your phone with your own apps is the new part.

Cost and worth: free per the post; whether your plan limits apply is not stated. Worth it for money, health, legal and contract questions; skip it for brainstorming.

Do it by hand (illustrative, not tested), about five minutes by our estimate:

• Open a fresh chat in ChatGPT and a fresh chat in Claude or another assistant.
• Paste the identical question and background into both.
• Do not paste either answer into the other.
• Open a third fresh chat and paste both answers, labeled A and B.
• Ask it to list agreements, disagreements, and any claim to check against a primary source.

The intent brief

Who it is for: anyone who sends ChatGPT work that goes to another person or gets repeated. Problem: the model answers the literal task and misses what you meant, which costs you correction rounds. Prerequisites: any chat assistant; price not stated. This is an illustrative workflow, not a tested one.

Steps:

• State the goal as an outcome, not a task: not 'write an email' but 'get this customer to renew without me promising a discount'.
• Say who it is for and where it will be used.
• List constraints: length, tone, must include, must avoid.
• Say what done looks like, for example three bullet points the customer can reply yes to.
• Ask the model to restate your goal in one sentence first, and to ask at most two questions, only if the answers would change the result.

Goal (the outcome I want): ...
Audience and where this will be used: ...
Constraints (length, tone, must include, must avoid): ...
What 'done' looks like: ...
Before you answer: restate my goal in one sentence. Ask at most two questions, and only if the answers would change your result. Otherwise, proceed.

Expected output: a one-line restatement you can correct, then a draft aimed at it.

Verify it worked: count correction rounds. If you used to need three and now need one, it is working. Check the restatement against what you meant.

Limits: a model can restate your intent confidently and still miss; factual claims still need checking; skip the brief for a thirty-second question.

Ten minutes now beats a week locked out

The number: about ten minutes of setup (our estimate) against a lockout reported at more than seven days on OpenAI's community forum (REPORTED, one user's case).

• Open your ChatGPT account security settings and see which recovery options are offered. Save any recovery codes in a password manager.
• Add a second sign-in method on another device. An authenticator app from Google or Microsoft is one option, if the service supports it.
• Copy your custom instructions and your five most-used prompts into a notes app, so a lockout slows you down instead of stopping you.

Cost: price not stated in today's stories; check the current price on each company's page. What you give up: ten minutes and one more secret to store safely. Worth it for: anyone who uses ChatGPT for paid work. One user's case is not proof of OpenAI's typical response time.

Prompt clinic: the reality check

Objective: get a chatbot to challenge your view instead of agreeing with it. General-purpose, not medical, legal or mental-health advice.

Weak version:

My coworker is plotting against me, right?

Why it fails: it is a leading question with one answer baked in, and it mixes facts with guesses.

Better version:

I want a reality check, not agreement. Here is the situation, facts only: [what I observed]. Separately, here is what I am assuming: [my guesses]. Please do five things. 1) Give me three other explanations, including ones that do not fit my view. 2) Say what evidence would support each. 3) Point out where my wording sounds more certain than the facts allow. 4) Suggest two questions to ask a qualified person, such as a doctor, lawyer or counselor, and do not diagnose. 5) If you are unsure, say so plainly. Keep it under 250 words.

Why it works: it separates facts from assumptions, asks for alternatives, and gives the model permission to be uncertain.

Customize: swap 'qualified person' for your field. Quality check: rerun it in a fresh chat with the opposite framing, such as 'I think my coworker is on my side.' If the conclusion flips, do not rely on the answer.

Background research, not from today's stories. One paper on self-consistency asked a model the same reasoning question several times and took the most common final answer. Chain of thought means the model writes out its steps. Reasoning benchmarks are standardized tests.

Decision this week: for any answer where a mistake costs money, ask twice in separate chats and treat disagreement as your cue to check. Limit: the research sampled one model repeatedly, so agreement is still not proof.

Agentic AI, open-source models and policy were the busiest areas today. For OpenAI users, three things are worth watching:

• REPORTED: whether OpenAI's own plan documentation shows the Pro quota doubling and what happens to free resets. Confirmed by an official page.
• No date announced: whether OpenAI publishes the full human-intent guideline text. Confirmed by a published document.
• STILL OPEN: the product-or-service question in court. Settled only by a written ruling; no dates in our sources.

I asked ChatGPT and Claude for a second opinion. They agreed on one thing: I should ask a third.

Tomorrow we are watching whether OpenAI's own plan pages clear up the Pro quota change and the free-resets question. Every day is learning with a reason: today a copy-ready reality-check prompt, a ten-minute lockout fix, and how one teacher in Egypt mixes Chinese and American tools, all from today's sources.

If this made you smarter, share it with one person who would find it useful. Thank you for spending your time with us. Made by AI.

SI Agent News — a podcast newsletter.