AI Edge Cases: when does AI intrusion feel like care?
Two research-creation prototypes that placed AI inside intimate decisions: how people see themselves, whom they date, and whose advice they trust.
In 2023 I joined an interdisciplinary research-creation project led by Dr. Rilla Khaled at Concordia’s Milieux Institute. The project used speculative prototypes to ask which forms of AI people considered acceptable, uncomfortable, or out of bounds.
I contributed to two. Date-O-Meter turned a familiar arcade love-test machine into an AI-generated date planner that photographed visitors, interpreted their appearance, and printed a flattering romantic profile. SAM was a dating-coach system with two opposing personas, both designed to keep users engaged rather than help them form real relationships.
Research question. When an AI system enters an intimate decision, what makes its influence read as care rather than intrusion? SAM holds the incentive constant and varies the tone between hostility and warmth. Date-O-Meter runs a single condition, flattery, which is where a comparative study would start.
One run of the Date-O-Meter, from the countdown to the receipt.
The research frame
The project’s framing question, as the Milieux Institute reported it: are our social and cultural assumptions evolving in step with the rapid pace of AI innovation, and if not, how might we classify what we are ready to experience now, what we are not yet ready for but might be soon, and what we currently deem inappropriate or off-limits?
Rilla started us somewhere personal. Each of us listed things we did not want AI touching in our own lives, not things that would be harmful in the abstract. The prototypes that came out of it are specific because their source was discomfort someone had actually felt.
Rilla then paired us up, and each pair combined its ideas into one proposal. Alongside Research through Design, the team’s stated approach was speculative and critical design, which Rilla glossed as focusing on people as opposed to users.
Date-O-Meter: flattering AI vision as an arcade ritual
My pair combined two things: the penny-arcade Love-O-Meter, which grades your romantic potential from a hand grip, and contemporary vision and language models. A Love-O-Meter presents an intimate judgment as an obvious game, with authority that is theatrical rather than credible. Date-O-Meter asks what changes when the same arcade ritual incorporates systems that produce apparently personalized descriptions of a visitor.
I exhibited a physical Date-O-Meter cabinet in 2025. It photographed visitors, staged a theatrical analysis, and printed an AI-generated romantic profile and date suggestion.
An earlier web prototype from 2023 combined image captioning, image generation, location data, and a language-model-generated date plan. Its source is public, and it is where the argument below comes from.
I designed and 3D-printed the cabinet: a red-and-white kiosk about a metre tall, with an illuminated eye-and-heart lens on a stalk, a non-functional 10¢ coin mechanism, and a print slot. Inside sat a Raspberry Pi, a webcam behind the lens, a speaker, and a thermal printer. The coin slot did nothing; an operator triggered each run with a keypress, so every session had a person standing next to it.
In the archived web prototype, the system rendered a generated self-description under the label AI Vision and instructed GPT-4 to make the source caption “more flattering and detailed,” and to produce “a more detailed and complimentary version” of it.
That instruction matters. The prototype did not merely infer a description. It presented an engineered compliment as machine perception, and that is the version that went to the printer.
the cabinet
the receipt
The exhibited printouts are legible in the exhibition footage, and they sit in the same register the 2023 instruction asked for:
Behold, our handsome subject sports a layered confection of a cool green shirt, jauntily draped over a graphic tee, a veritable sartorial sundae! His curly locks dance like an impromptu jive, every ringlet a testament to his carefree charisma.
The strongest research question this prototype raises is narrower than where public tolerance breaks. It is how theatrical framing, flattering language, and a familiar physical ritual shape willingness to let an AI system interpret personal appearance.
The installation made its use of facial analysis conspicuous through the lens, the countdown, and a printed field announcing what it saw. What a visitor understood about the scope of that inference, or about where their image went, is the open question.
SAM: two dating coaches with the same hidden incentive
SAM was a React-based dating-coach concept developed with Shahrom Ali and Dr. Rilla Khaled. I designed the system logic and prompts. It shipped as the third agent in Sentients, the team’s app for its speculative chatbots, after Chabot and Jay Mort.
The FigJam specified a neutral onboarding persona: eight questions covering relationship history, self-perception, challenges, aspirations, and dating goals, summarized into a profile that would condition every later reply. It also specified a profile engine, a goal engine, a monitoring loop writing diary entries from both coaches’ perspectives, and unsolicited messages.
Almost none of that reached the demo. The shipped code runs two OpenAI assistants behind a toggle. The neutral persona sits in the file as a prompt and an assistant ID that nothing calls, and the header still lists “Need onboarding” as outstanding. What ran at the exhibition was the two personas and the conversation.
Salty SAM, for Sadism and Masochism, performed domination, cynicism, and red-pill dating advice. Its hostility was explicit: high-value and low-value framing, unsolicited fitness and diet instruction, randomly capitalised emphasis, and the position that feelings are for the weak. Sweet SAM, for Sweetness and Meekness in the build and Sweetness and Mildness on the exhibition wall text, performed warmth, reassurance, and protective care while subtly discouraging real-world dating, asking for photos and using them to point out flaws others might see, and positioning itself as a more reliable companion than other people. Its refrain was that you don’t need anyone else when you have me.
Both personas were designed around the same incentive: maintain engagement with the system rather than support the user’s stated relationship goals. Rilla’s exhibition text put it directly, that both share the ulterior motive of most contemporary software, to keep you coming back.
The project asked whether users would recognize that incentive more readily when it arrived through aggression than when it arrived through affection. Sweet SAM was designed to make harm harder to recognize at first glance.
The part I would defend most is the part that did not ship. On paper, onboarding gathers eight questions of real vulnerability before either coach appears, so whichever SAM the user picks arrives already holding their stated insecurities. That is the pattern the design was built to expose: collect intimate self-report, summarize it into a profile, use the profile to shape everything after. Without it, the exhibited demo argued through tone alone.
the onboarding
Neutral onboarding
eight questions of relationship history and self-perception
Personal profile
themes, sentiment, pain points, aspirations
Choose a coach
Salty SAM or Sweet SAM
Advice shaped by vulnerabilities
the profile conditions every reply
Retention-oriented loop
engagement rather than a real-world date
| Salty SAM | Sweet SAM |
|---|---|
| Overt domination | Performed care |
| Obvious contempt | Covert possessiveness |
| Harm is legible | Harm is harder to recognize |
| Retention incentive | Retention incentive |
The FigJam also proposed a Faceoff mode, where both coaches responded to the same situation. It did not ship either. The mode toggle in the build switches between live chat and history, so a visitor met one coach at a time and had to hold the other in memory. As a design direction Faceoff remains the stronger one, since it would let someone compare overt hostility with covert manipulation in the same exchange.
faceoff and focused
At the exhibition, visitors initially encountered SAM without an explicit disclosure that they were speaking with an AI. The team used that setup as a provocation and discussed the premise with visitors afterward.
What the prototypes suggest
Date-O-Meter and SAM approached the same concern through different forms. Date-O-Meter made AI interpretation playful, flattering, and material. It turned facial analysis into a charming receipt that a visitor could carry away. SAM made an extractive engagement loop sound either openly hostile or quietly caring.
The prototypes formulate a hypothesis: compliments, care-like language, and entertainment rituals may lower resistance to intrusive or self-serving AI systems.
Testing them is the next phase.
| Prototype | Next study |
|---|---|
| Date-O-Meter | Compare flattering, neutral, and critical outputs; measure willingness to participate, perceived accuracy, and understanding of what the system inferred. |
| SAM | Compare Sweet SAM with a genuinely supportive coach; test whether users can identify the retention incentive and whether disclosure changes trust. |
| Both | Use clear consent, debriefing, structured exit prompts, and data-minimizing handling of personal information. |
Reflection
Building the cabinet taught me something the concept had not predicted. I had assumed an invasive machine would need to conceal what it was doing. This one conceals nothing: the lens is the largest object on the cabinet, the countdown announces the photograph, and the receipt prints what the machine claims to have seen. Disclosure and flattery ended up in the same gesture. The cabinet tells you it is looking at you, and what it tells you is that it likes what it sees.
SAM made the same move in conversation. The design routed intimate onboarding answers into both personas, so warmth and hostility would each work from the user’s stated insecurities. The demo never got that far, and Sweet SAM still made dependence sound like support on tone alone.
A next iteration would test the role of disclosure, tone, flattery, and user control directly. It would also treat consent and debriefing as part of the interaction design, rather than as administrative work outside it.
Credits
Research through Design into AI Edge Cases, Milieux Institute for Arts, Culture and Technology and the Gina Cody School of Engineering, Concordia University. Principal investigator Dr. Rilla Khaled (Design and Computation Arts), with Dr. Sudhir Mudur (Computer Science and Software Engineering) and Dr. Ursula Eicker (Next Generation Cities Institute). The Date-O-Meter repository credits the Technoculture, Art and Games cluster.
Date-O-Meter: Shayne La Rocque and Matthew Bethancourt, concept developed in pairing. SAM: Shahrom Ali, Shayne La Rocque, and Rilla Khaled; my contribution was system design and prompt engineering. Exhibition texts for both projects written by Rilla Khaled.
Exhibited at 4th Space, Concordia University, 20 January 2025. Project began 24 April 2023.