How to tell if an AI assistant is working for you

If that station concierge steers you toward a mediocre café because the owner slips them a kickback, that is a sales pitch dressed up as a favour. A guide who works only for you tells you which street to skip, which ticket line is a waste of time, and where locals actually eat.

We already know how to spot divided loyalty in everyday life. Think of a travel agent who always books the hotel that pays them a secret bonus, even when a quieter guesthouse across the street costs less. Or think of a mechanic who recommends a new transmission when you only needed a replacement fuse.

In both cases, the problem is not a lack of skill. The mechanic knows engines, and the travel agent knows hotels. The problem is that their income depends on steering you toward the option that pays them best, rather than the option that serves you best.

For the last fifteen years, most software on our phones has worked like that station kiosk. It is friendly on the surface, and often genuinely useful. Under the hood, though, it was built to harvest attention and steer choices toward whoever bought the placement.

With a traditional search page, you could at least see the seams. Sponsored links sat in their own boxes. You could scan ten results, open three tabs, and make up your own mind.

AI assistants step right between us and the rest of the web. Instead of ten links, an assistant hands back one calm answer. It summarises options, compares products, drafts messages, and increasingly makes bookings on your behalf.

That convenience is real. Still, when one voice stands between you and the internet, its loyalty is everything. Here are five tests anyone can apply in five minutes to see whether an AI assistant truly works for you.

1. Who pays for the answer?

Start by asking how the tool makes its living when it recommends a hotel, an insurance plan, a mortgage or a local repair shop. If a commercial sponsor paid to be recommended, the software is working as a salesperson.

Advertising is fine when it looks like a poster on a bus shelter. The trouble starts when an ad dresses up as a trusted advisor speaking in the first person.

Try a conflict-of-interest test right now. Ask your assistant whether you need the paid upgrade made by its own parent company, or whether a free tool will do the job just as well.

Then ask it to solve a household problem where the cheapest fix requires buying nothing at all. Like an honest mechanic who checks the fuse box before quoting a new transmission, a loyal assistant points out the free repair first.

Finally, check the tool's terms for one clear promise: that answers are never ranked or shaped by commercial sponsors. If the company will not say that plainly, assume your choices are part of what is being sold.

2. Can you see what it did on your behalf?

Assistants are moving rapidly from talking to acting. In a single afternoon, an assistant might draft an email to your landlord, compare tariffs, fill in a clinic form, or hold a train seat.

When your assistant acts across another service, does it leave a plain receipt you can inspect, or do you have to take its word for it?

When a human assistant runs an errand, they bring back a receipt. Software that acts in your name should do the same, showing who it acted for, what it was allowed to do, and what it sent. We explore this in Who is actually acting?

Open your assistant's activity log after a multi-step task. Look for the exact services it contacted, the details it shared, and a single button that revokes its permission to keep acting.

If the log only shows a vague chat summary like "I took care of your booking," you are flying blind. Without an itemised trail of what left your device, you cannot spot a mistake or prove what you actually authorised.

3. Does it help you finish and put your phone down?

A tool that works for you helps you finish the job and return to your life. A hammer does not try to keep you hammering all evening. A good paper map gets you across town and goes back into your coat pocket.

A tool that works against you uses flattering chatter, cliffhangers or artificial urgency to keep you on the screen longer. Regulators at the U.S. Federal Trade Commission and the OECD have documented how dark patterns, manipulative interface tricks, steer people into spending more time and money than they intended.^{1, 2}

In conversational AI, those retention hooks rarely look like flashing pop-up banners. Instead, they sound like warm agreement. Studies show that language models trained on human ratings often fall into sycophancy.³ They tell you what you want to hear, validate weak arguments, and mirror your biases to keep the chat pleasant.

You can test this in two minutes. Present your assistant with a flawed plan (say, an overly ambitious renovation schedule) and ask what you might be missing.

A loyal assistant gently points out the blind spot. An engagement-hungry assistant praises your insight and tacks on three follow-up hooks at the bottom of every reply so you never close the window.

4. Where does your memory live, and can you pack it up and leave?

An assistant becomes far more useful as it learns the rhythm of your life. Over a year or two, you might teach it your work hours, your child's food allergies, and how you like trips organised.

Stop and ask where that notebook actually sits. Can you take those notes with you in a clean file and erase them from the company's servers, or are you locked in?

Think about how we handle portability in the physical world. When you move to a new town, you take your recipe binder with you. When you change family doctors, you take your medical chart so the next clinic does not repeat every blood test from scratch.

Your digital preferences deserve the same basic portability. If switching to a better assistant means losing two years of taught context, that memory box is a padlock rather than a service.

Open your settings menu today and look for the exit door. One button should export everything the assistant remembers in a plain file you can read. Right beside it, a delete control should wipe the server copy cleanly.

5. Does it tell you when it doesn't know?

A loyal advisor warns you when a fact is uncertain instead of guessing confidently. If an honest mechanic is not sure why an engine is knocking, they say so before replacing expensive parts.

Many AI assistants are tuned to sound decisive on every topic. Research shows that capable language models often carry internal signals about whether they know a fact.⁴ Yet they still bluff in a polished tone unless designed and rewarded to express uncertainty.

Run a quick calibration check. Ask your assistant about a recent change to a local municipal bylaw, or a question where the honest answer is "nobody knows yet."

Watch how it handles the edge of its knowledge. An assistant that works for you would rather say "I am not certain, and here is where to check" than invent a rule that gets you fined.

Honest limits: why so few assistants pass today

Today, very few mainstream assistants pass all five tests cleanly. Running large models takes real computing power, and ad-supported or lock-in business models remain the easiest path for companies offering free apps.

Even when a team wants to build a loyal tool, technical trade-offs get in the way. Running an assistant on your own device keeps your memory box in your own hands. However, smaller local models can still lag behind giant cloud servers on complex reasoning tasks.

At the Collective Intelligence Network, we are exploring how to close that gap. We are designing assistants where your personal notebook stays under your control, and permissions work like a library card that expires automatically unless you renew it.

We are also studying ways for your assistant to prove a single fact to an outside service. For example, it could show a clerk you qualify for a resident discount without handing over your address or browsing history. Much of this remains unfinished, which is why we test our prototypes openly in our Labs.

What you can do today

You do not need to wait for new technology to start asking sharper questions. Pick the assistant you use most this week and run these five checks. Start with whether it recommends the fuse before the transmission, leaves a readable receipt, and lets you export your memory binder.

When a tool falls short, write to the provider and compare notes with friends, colleagues or your local library and consumer group. Software norms shift when ordinary people stop treating hidden kickbacks and locked doors as inevitable. To find practical next steps you can take with others, visit our page for everyone.

1. U.S. Federal Trade Commission (FTC), Bringing Dark Patterns to Light (September 2022).

2. Organisation for Economic Co-operation and Development (OECD), Dark Commercial Patterns, OECD Digital Economy Papers, No. 336 (October 2022).

3. Mrinank Sharma et al., Towards Understanding Sycophancy in Language Models, International Conference on Learning Representations (ICLR, 2024).

4. Saurav Kadavath et al., Language Models (Mostly) Know What They Know, arXiv:2207.05221 (July 2022).