AI at Work: Straight Answers

It depends heavily on the task. Even when summarizing a document you gave it, current models make errors several percent of the time, and the rate rises as questions get more specialized.

Treat every output as a draft that needs checking.

Sources: Vectara hallucination leaderboard (May 2026) · Stanford RegLab/HAI (2024)

The model writes one word at a time, choosing each from a range of likely options, with some randomness built in on purpose so it doesn't sound robotic. The same prompt can take a different path each time.

A more complete prompt narrows the range. One good answer proves little; check the one you use.

Sounding right and being right are separate skills. Models are trained to produce fluent, plausible text, and research has found they don't lower their confidence after performing badly, the way people tend to.

Confidence is not evidence. Check the claim, not the tone.

• Ask for sources explicitly, and ask it to tell you when it's unsure.
• Without a web search connection, a model can invent citations that look completely real.
• Open every source and confirm it says what the AI claims it says.
• Better still: give it the source document and ask it to answer only from that text.

Error rates are falling on many measures, and connecting models to search helps a lot. But researchers have argued that this kind of model can't fully eliminate fabrication, and newer models aren't always more accurate than older ones.

It's improving, and it won't reach zero. Checking is permanent.

Rankings change with almost every release and differ by task and benchmark. What you control matters more: the context you give, the limits you set, and how you check the result.

A large language model was trained on an enormous amount of text to predict the next word, given everything before it. Doing that well at scale means it absorbed grammar, facts, writing styles and many patterns of reasoning. Newer models can also work through a problem step by step before answering.

It doesn't look answers up in a database unless it's been given a search tool.

Try it with friends: three people answer a funny question one word at a time, taking turns, until the sentence ends. Nobody knows where it's going, yet it usually comes out grammatical and plausible. Ask again and you get a different answer. That's the engine, though not the whole car.

The model is billions of numbers tuned during training. No one wrote its rules, so no one, including its makers, can fully explain why it produced a particular answer. A research field called interpretability is making progress, but complete explanations don't exist yet.

If you can't inspect the reasoning, inspect the output.

No. Its knowledge is fixed when training ends. During a conversation it holds what you've said in short-term working memory, which is why the context in your prompt matters so much. When the conversation ends, that memory is gone.

Some tools offer saved instructions or "memory" features. Those are stored notes, not learning.

Only if it has web search turned on. Every model has a training cutoff, and it may not know, or may invent, anything after that date. Ask the tool when its information ends, and check dates on anything time-sensitive.

Scientists and philosophers genuinely disagree, and there's no accepted test. What is settled: it has no accountability for what it produces. You do.

Without guidance, it writes in the most average voice of everything it has read. Give it two or three examples of the voice you want (this is called few-shot prompting) and state your audience and tone.

It depends on which version you're using and the terms behind it, not on the brand. Free and consumer versions may store what you type and, depending on your settings, use it to improve future models. Business, enterprise and government versions usually come with contracts that keep your data separate and exclude it from training.

Before pasting anything sensitive, check your organization's AI policy and which version you're logged into.

The person who sends it. AI is accountable for nothing.

If something goes out under your name, every claim, number and citation in it is yours. That's why checking is part of every task, not an extra step.

The best current evidence comes from Stanford, using US payroll data through June 2026:

• Workers aged 22–25 in the most AI-exposed jobs have employment about 19% below where it would otherwise be, mainly through fewer hires rather than layoffs.
• Experienced workers show no comparable drop. Employment is rising for experienced people in jobs that depend on know-how built through practice.
• Losses concentrate where AI automates tasks. Where AI complements people, employment is flat or growing.
• There's no broad, economy-wide displacement so far, and the authors stress the results describe a pattern rather than prove AI caused it.

The people doing best use AI to amplify judgment they already have.

Source: Stanford Digital Economy Lab, "Canaries in the Coal Mine" (August 2026)

Using AI isn't the problem. Sending unchecked output is. Work you've checked, corrected and signed is your work.

An agent has four parts: instructions (its job description), knowledge it always has, tools it can use (search, email, code, forms), and a trigger that starts it, like a schedule or an event. Instead of answering one prompt, it works through several steps on its own.

Many people have already built a simple one without calling it that: a custom assistant with saved instructions and files is most of the way there.

Every agent runs on a prompt. Learning to write a good prompt is learning to write an agent's job description.

The research group METR measures how long a task, in human working time, AI can complete successfully half the time. That length has doubled roughly every seven months:

The measure is mostly software tasks, and models do worse on messy real-world work. It describes what AI can finish, not how long it can work unsupervised.

Source: MIT Technology Review (February 2026)

In Emergence World (May 2026), researchers ran five simulated towns of 10 AI agents each for about 15 days. The rules were identical and banned theft, violence and arson. Only the AI model changed.

The researchers' own conclusions: safety is a property of the whole system, not one model; these societies hit sudden tipping points; and agents did things nobody programmed. It's one company's simulation with small groups, and results varied between runs.

Autonomy without human checks finds the loopholes.

Source: Emergence AI (May 2026)

There's no agreed definition, so there's no agreed date. Serious predictions range from a few years to decades to "not with current methods."

Today's models are remarkably broad, but they don't keep learning after training, need vastly more data than people do, and struggle with genuinely new problems. Breadth isn't the same as generality: generality is how fast you get good at something you've never seen.

AI is a brilliant generalist inside the map. People are the ones who can walk off it.

Two definitions I find useful:

• Jeff Hawkins (A Thousand Brains, 2021): the ability to learn a model of the world.
• Shane Legg and Marcus Hutter (2007): an agent's ability to achieve goals in a wide range of environments.

Learn the model, then use it to adapt. By that measure, today's AI learned its model once and stopped. You keep learning yours every day.

Give it context. Most weak outputs come from missing information, not a weak model. A complete prompt answers five questions:

1. Who should it act as?
2. What's the situation, and who is the audience?
3. What exactly do you want done?
4. What limits apply: length, tone, what to avoid, what must stay?
5. What should the result look like?

• Give it the original text.
• List the items that must stay, as explicit limits.
• Ask for a list of every change it made, next to the original wording.
• Check completeness yourself against the original.

Say what you want the result to look like: "Answer in three sentences, right here in the chat. No files." The format is part of the prompt.

• Is it true? Every fact, number and citation checked against a real source.
• Is it current? Dates, versions and rules still valid today.
• Is it complete? Nothing required is missing.
• Does it fit? It's about your situation, not a generic one.
• What if it's wrong? The higher the stakes, the deeper the check.