How to Trick AI: The Battle for Personality Privacy in the Age of Chatbots

TL;DR

Can you hide your soul from an algorithm? This CHI '20 study explores how users attempt to "trick" personality-assessment AI. While users can shift their AI-generated scores by nearly 10% through strategic lying and stylistic changes, the sheer mental effort required makes manual privacy protection nearly impossible in the long run.

Context: The Invisible Mirror

We are increasingly living in a world where our word choices, response latencies, and even use of punctuation serve as a "digital DNA." Automatic Personality Assessment (APA) systems, like the Juji chatbot used in this study, can analyze these cues to map us against the Big Five (OCEAN) traits: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism.

The danger? These profiles are often used for Psychological Targeting—from manipulative political ads (á la Cambridge Analytica) to opaque job screening processes.

The "Tricking" Framework: 41 Strategies of Deception

The researchers identified a massive taxonomy of 41 strategies used by participants to mislead the AI. These can be categorized into four primary domains:

1. Linguistic Syntax: Manipulating grammar, casing, and punctuation (e.g., using no periods to appear less "Conscientious").
2. Word Choice & Style: Inserting specific keywords, using "filler" words, or lengthening sentences to mask true intent.
3. Content Semantics: Reporting "opposite" behaviors or lying about personal interests.
4. Conversation Dynamics: Deliberately slowing down or speeding up response times (latency) to appear more introverted or neurotic.

Figure: The gap between what users identify as a cue and what they can actually manipulate in real-time.

Methodology: The Interaction Experiment

The researchers employed a clever two-phase design:

• Phase 1 (Baseline): Participants chatted for 45 minutes in a realistic customer service scenario (booking a holiday, buying a backpack). The goal was to capture their "True" digital footprint.
• Phase 2 (The Trick): Participants were told the AI was harvesting their data and offered a financial incentive to make their second profile differ as much as possible from the first.

Figure: The experimental workflow from natural interaction to strategic manipulation.

Why Tricking AI is Harder Than You Think

The results revealed a fascinating psychological barrier. While participants successfully shifted their scores by roughly 8-10%, they found the process "exhausting and tedious."

• The Implicit Leakage: You might remember to avoid "swear words" to appear more Agreeable, but can you remember to use fewer personal pronouns ($I,me,my$)? Most can't.
• Cognitive Load: Trying to maintain a "persona" while simultaneously solving a task (like booking a flight) leads to "fatigue effects."
• Accuracy Paradox: Interestingly, the study found that the AI itself wasn't perfectly accurate compared to traditional BFI-2 questionnaires, yet users still tended to "over-trust" the AI's assessment of them, sometimes questioning their own self-image when the AI disagreed.

Experimental Quantitative Results

The absolute difference achieved across the traits was significant but not transformative:

Deep Insight: The Future of Privacy-Protective Design

The authors conclude that self-protection is not enough. If we cannot effectively "fake it" without mental burnout, the burden of protection must shift to designers and policymakers.

Three provocative directions for future tech:

1. Machine Non-Readable Content: Developing systems that render text as distorted images for machines but clear text for humans (like "Text CAPTCHAs" for every post).
2. Privacy Feedback Tools: Real-time writing assistants that warn you: "This sentence makes you sound highly neurotic; consider this alternative to protect your profile."
3. Legislative Enforcement: Moving beyond "endless license agreements" to mandatory opt-in consent for any form of automatic personality inference.

Conclusion

This paper serves as a wake-up call. Personality assessment is no longer the domain of psychologists with clipboards; it is a silent, background process of the modern web. While we can "trick" the AI to a point, our distinct patterns of behavior are incredibly resilient. True privacy in the AI era may require not just better strategies, but a fundamental redesign of how we interact with machines.