How leaders can use psychological targeting for good

Executive overview

Algorithms now build detailed psychological profiles from digital behaviour — social media language, GPS data, photos — and use them to influence decisions at scale. Most organisations collect this data to extract value from users; the same tools can instead serve users' genuine interests.

Two practical moves for leaders: involve users transparently so targeting is consensual and accurate, and adopt privacy-preserving architectures (federated learning) so data never needs to be hoarded in the first place.

The same psychological targeting that sells people things they don't need can, with consent, help them save money, access mental health support, and make better decisions.

How algorithms read psychology from digital traces

• Language patterns on social media reliably predict income, personality, and emotional state.
• Low-income individuals use more present-tense language and more self-references — markers of financial stress, not deliberate signals.
• Extroverts and introverts show distinct word clouds: social activity vs. solitary interests.
• Facial features and grooming choices in photos predict personality traits with above-chance accuracy, even when camera angle and styling are controlled.
• Algorithms detect these patterns by counting relationships across millions of data points — not a black box, but scale humans can't match.

The dark side: manipulation and unavoidable tracking

• Marketers already use psychological profiles to increase purchase likelihood — the same lever pulls in both directions.
• Facial recognition combined with ubiquitous cameras means opting out of tracking is practically impossible.
• Pseudoscientific physiognomy was historically abused; modern AI can now do what it could not — find real statistical relationships between facial features and psychology.
• Passive smartphone data (GPS, call frequency, physical activity) can detect early signs of depression before the person recognises them or reaches out.

Using targeting for good: the Safer Life example

• Safer Life, a fintech for low-income users, ran a "Race to 100" challenge — save $100 in a month, starting from under $100 in savings.
• Psychologically tailored messages were tested against Safer Life's own optimised gold-standard messages.
• Agreeable users (high social orientation) received messages framing savings as protecting loved ones, not accumulating wealth.
• Result: 60% increase in users hitting their savings goal compared to the gold standard.
• The same targeting infrastructure that drives spending can be redirected to support financial resilience.

Mental health: tracking and treatment

• Depression reduces the impulse to reach out — passive monitoring can flag deviations from a baseline before a crisis deepens.
• Early-warning signals: reduced physical activity, fewer calls, less time away from home, changes in emotional language online.
• Users can nominate trusted contacts to be alerted when deviations appear, enabling a human response before the person is deep in crisis.
• WHO data: roughly 13 professional therapists per 100,000 people seeking care — chatbots fill a real gap, especially at 3am.
• AI therapy is not a replacement for professional care, but meaningfully better than no care at all.

What leaders should do

• Involve users explicitly. Predictive models make mistakes at the individual level; consent conversations surface errors and build trust. Hilton's traveller-profile project showed users their predictions and let them correct them — turning data collection into a value proposition.
• Stop hoarding data by default. The assumption that more data is always better is no longer true.
• Adopt federated learning where possible. The model goes to the user's device, learns locally, and returns only intelligence — not raw data. Apple and Google both back open-source federated frameworks.
• Run the evil Steve test. Apple's design practice: assume a future leader with opposite values inherits every system you build today. If that scenario makes your data architecture dangerous, redesign it now.

Why putting responsibility on users doesn't work

• Privacy regulations in Europe and California rely on transparency and user control — correct in principle, unworkable in practice.
• Even researchers who spend years studying this topic routinely click "accept all" because managing permissions is a full-time job.
• Genuine protection requires privacy by design at the regulatory and architectural level, not informed consent forms.
• Federated learning eliminates the trade-off between personalisation and privacy — users get the service without surrendering the data.