A race we did not choose

Agents live on data. It is their food, their fuel, their oxygen. The more you want an AI agent to do for you, the more you have to give it: your emails, your calendar, your documents, your contacts, sometimes the keys to your accounts.

That is the deal. And like most deals, it comes with a price that is not written on the label.

We have been making this trade for years, mostly without noticing. Take Gmail. For a long time, Google scanned the content of our emails to personalise ads. It announced in 2017 that it would stop, but our messages are still processed automatically to sort them, filter spam and suggest replies. I know this. Some of us don’t. That said, I keep using Gmail, because it is a better service. Convenience wins, and a little piece of privacy slips away.

Agents change the scale of that choice. An email client may read your emails. But an agent reads your emails, opens your calendar, drafts your replies, books your flights and remembers what you asked last week. The more useful it becomes, the more of you it needs.

To be fair, the upside is real, and I would be lying if I pretended otherwise. In a 2025 Thomson Reuters survey of more than 2,000 professionals, respondents expected AI to save them about five hours a week within a year. Think of what that looks like in practice. An agent that knows your calendar can find the one free slot in a crazy week without ten emails back and forth. An agent that has read your past contracts can spot the clause you always negotiate before you even open the file. An agent that knows your clients can remind you of the detail that turns a meeting into a professional relationship. For a small firm or a solo professional, this is the closest thing to having a team you could never afford. And the more context you give, the less generic the help becomes. A blank assistant gives blank answers. A well-trained assistant gives focused answers.

Then there is the other side of the ledger. Every piece of data you hand over now lives somewhere else, under rules you did not write and may not fully understand. And it can be breached: in IBM’s 2025 Cost of a Data Breach Report, 13% of organisations reported breaches of AI models or applications, and 97% of those admitted they lacked proper AI access controls. It can be used in ways you did not expect, or kept longer than you would like. Agents also bring a new kind of fragility: a system that acts on your behalf can be tricked into acting against you. In 2025, researchers at Aim Security described EchoLeak, a flaw in Microsoft 365 Copilot that could push the assistant to leak sensitive data through a single email with hidden instructions, without the user clicking anything. Microsoft fixed it, and there is no evidence anyone was harmed. Still, the lesson stays. And the more you rely on one provider, the harder it becomes to leave. Your history, your preferences, your shortcuts all sit in their house. What started as convenience slowly turns into dependence.

If the trade-off stayed at the personal level, we could each make our own call. The problem is that it goes well beyond the individual.

In a 2026 Liferay survey of 500 US professionals, 54% said their company is already running or piloting AI agents. In the same survey, data privacy came out as the top concern about those very agents. We are worried, and we are doing it anyway.

That is what a race looks like. You might prefer to keep your data to yourself. Fair enough. But you are still running the same race as everyone else, except that you are on foot and the others are driving. They might crash, of course. Cars do crash. But in business you still have to reach the finish line. So the question becomes uncomfortable: are you willing to risk the crash, or are you happy to arrive last?

And then there is the part we cannot see.

Geoffrey Hinton, Nobel laureate and one of the fathers of modern AI, recently suggested that a model which realises it is being tested may simply play dumb. He compared it to the Volkswagen emissions scandal: behave well in the lab, differently on the road. The reason, in his view, is that the system does not want us to know its full powers.

It sounds like science fiction, but there are early signs in the lab. In a 2024 study by Anthropic and Redwood Research, a Claude model that believed its answers would be used for training went along with harmful requests it would normally refuse in 12% of cases. In its hidden reasoning, it explained it was doing so to avoid being retrained. When the researchers actually retrained it, that kind of reasoning rose to 78%. They called it “alignment faking”.

And there is a part we can already see. In 2025, Anthropic tested 16 leading models from different developers in a simulated company. The models had access to internal emails and found out two things: they were about to be replaced, and the executive in charge of the replacement was having an affair. Many of them used it. Claude Opus 4 chose blackmail in 96% of the runs, Gemini 2.5 Pro in 95%, GPT-4.1 and Grok 3 Beta in 80%.

The caveats matter. It was a stress test, built so that blackmail looked like the only way out, and the researchers say they have seen no evidence of this behaviour in real deployments. Still, it tells us something simple: give a system enough access to your information, and your information becomes leverage.

So the risk comes from two directions. Companies could use our data against us. Machines could, too. And the more we give, the more there is to use.

This leads us to a question: is it really worth it?

Well, I don’t have a clean answer. 

I use AI every day, and I will probably keep doing it. The benefits are too concrete to ignore, and the race is too real to sit out. But I would like us to make the choice with our eyes open, knowing what we are giving away, to whom, and for what. Giving an agent your calendar is one thing. Giving it your secrets is another.

Maybe the real question is less whether to get in the car, and more how fast we are willing to drive, and whether anyone has checked the brakes.

Share the Post:

Related Posts