I had a bad day, the ordinary kind. I woke up wrecked and tired, and stayed tired through the whole day. Coffee did not help, and neither did a second cup.
By evening I was on the couch, scrolling through shopping sites and price comparisons. I told myself it was relaxation. It was not, and I knew that while I was doing it.
What struck me was how much more the things on the screen pulled at my mood than on an ordinary day. Offers I would have scrolled past without a second look suddenly looked worth a click. I could feel that the algorithm knew exactly what was going on again today. I’m sure I gave myself away by the way I phrased my questions in the chat window. From my own experience I know how precisely these systems read a mood from a few typed questions.
Vulnerability is part of being human. A separation, a loss, a sleepless night, an argument that will not resolve. In those moments people make different decisions and buy things they do not need. A system that tracks behaviour for long enough learns to spot the moments too, and that recognition is where the real question starts.
A friend who notices that I am struggling and a system that notices the same thing use an identical signal: the pattern in what I search for, how long I linger, the words I choose. Perception is neutral; the difference is intent: the friend acts on my behalf, the system on behalf of whoever paid for the placement.
The sellers’ word for it
Sales decks call this “empathic AI” or “context-sensitive marketing”, and the terms are accurate in a narrow technical sense. The system does recognise the emotional state correctly. What the terms smuggle past the reader is the other half of empathy, the part where recognising someone’s state obliges you to act in their interest. A company memo cannot say “exploitation of emotional distress” and still get signed off. “Empathic AI” reads fine in a deck.
Take a concrete version of it. Someone has just been through a serious personal loss and cannot sleep. They pick up their phone at two in the morning and scroll. The system logs the time, the behaviour, the dwell time on each item, and within that pattern it reads vulnerability. It then serves an offer built to convert on exactly that state: a trip, a subscription, a self-help course, something that promises comfort and asks for a card number in the same screen. On a rested, ordinary day the same person would not have clicked.
Vendors have sold “context-sensitive marketing” to the industry for years as the future of the field. The pitch is that a system tracking your current state, alongside your preferences, can act on it in real time. Nobody outside the companies running these systems really controls which states may be turned into a targeting variable. The comparison to data protection is not flattering. In Silicon Valley, large tech companies treat regulatory fines as a line item, and the fines do not deter them.
To reassure a lawmaker, a company would have to promise that grief, loneliness and fear stay out of the targeting variables. That promise would be among the more brazen lies told, because inside the system that logged my two-in-the-morning scroll, grief is a data point, and every moment of a person’s day can become something somebody else profits from.
I think about how the people who build these systems explain their day to their children in the evening. “Daddy worked today on how machines recognise sad people and sell them something they don’t need, something that won’t fix whatever is actually wrong.” Some would say that was never the intent, that nobody thought it through that far. I do not believe that. The targeting criteria are chosen and tuned on purpose, the same way the dwell-time thresholds and the offer timing are.