The Mathematical Soulmate: AI in Personal Relationships and Matchmaking

AI can reduce search friction, rank profiles, surface shared interests, detect some abusive behavior, and help people express preferences. It cannot calculate a soulmate, infer consent, read lasting compatibility from facial “micro-expressions,” or establish relationship success from genetic data.

A responsible matchmaking system is therefore not a compatibility oracle. It is a constrained recommendation and safety service operating on intimate, incomplete, and changing information. Its quality depends on whether it expands meaningful choice without manipulating attention, exposing sensitive data, or turning popularity into destiny.

1. Define the relationship outcome honestly

“More matches” can mean more swipes, mutual likes, replies, conversations, in-person meetings, second dates, or durable relationships. These outcomes are not interchangeable. Optimizing session time may conflict with helping a person leave the platform after a good match.

Create an outcome hierarchy with explicit limits:

• healthy discovery: relevant profiles seen without excessive repetition;
• mutual interest: consent from both participants, not unilateral targeting;
• conversation quality: reciprocal exchange without harassment or fraud;
• user-defined progress: a meeting, friendship, partnership, or another declared goal;
• safety: low rates of impersonation, coercion, abuse, and unwanted recontact;
• autonomy: easy control over filters, visibility, explanations, pauses, and deletion.

Do not claim long-term compatibility when labels stop at a reply or first meeting. A proxy can help product measurement, but it must retain its name.

2. Learn from matching research without overclaiming

An original speed-dating study, “Is Romantic Desire Predictable?”, applied machine learning to more than 100 self-reported traits and preferences. The models predicted some general tendency to desire or be desired, but not the unique relationship-specific variance before two people met. That result came from two speed-dating studies and does not prove that every future system must fail; it does refute casual claims that profile questionnaires already reveal pair-specific chemistry.

Another observational study, “Aspirational pursuit of mates in online dating markets”, analyzed messaging in four US cities on one heterosexual dating service. It found patterned pursuit and response behavior within that dataset. It did not establish a universal hierarchy of human worth, causal rules for attraction, or results applicable to every culture, orientation, product, and period.

The practical lesson is modest: matching systems can model observed platform behavior, but observed behavior is shaped by interface, exposure, competition, norms, and who joined. It is not a stable biological truth.

3. Build a consented data map

Inventory every field and inference: profile text, photos, age, location, orientation, preferences, messages, likes, blocks, reports, device signals, verification records, and model-derived attributes. For each, record purpose, lawful basis where applicable, retention, access, sensitivity, and deletion path.

Collect the least information needed. Do not infer race, religion, health, disability, income, or sexuality from photos and language merely because a model can. Do not use private messages to train general-purpose models without a clear, specific choice. Location should be coarsened for discovery and separated from fraud or safety systems.

The FTC’s March 2026 action involving OkCupid and Match Group alleged that personal information, including photos and location, was shared with an unrelated third party contrary to privacy promises. That settlement concerns specific alleged conduct under US consumer-protection law; it is not a complete global privacy code. Operationally, it demonstrates why product promises, vendor access, and actual data flows must agree.

4. Separate candidate generation, ranking, and safety

A production system usually has several layers:

1. eligibility rules remove profiles that conflict with explicit age, location, orientation, or availability constraints;
2. candidate generation retrieves a manageable set using stated preferences and behavioral signals;
3. ranking estimates relevance or mutual-interest probability;
4. diversification prevents one look, neighborhood, or popularity tier from dominating;
5. safety services screen verification risk, spam, harassment, and policy violations;
6. the interface explains key controls and lets people decline or reset personalization.

Safety scores should not be silently converted into attractiveness scores. Likewise, moderation evidence must not leak into public labels that expose reporters or alleged victims. Keep recommendation, integrity, and enforcement models separate, with documented data-sharing boundaries.

For higher-impact removals or restrictions, use the review patterns in AI content moderation and digital trust. Human review should receive relevant evidence, policy, uncertainty, and an appeal path rather than a model verdict alone.

5. Design training and evaluation around people and time

Random interaction splits leak the same users into training and test data. Hold out people, new-user cohorts, cities, time periods, and product versions. Evaluate cold start separately from established users, and do not let post-match behavior become a feature for predicting the match that preceded it.

Offline metrics should include:

• ranking quality for mutually eligible candidates;
• calibration of mutual-interest estimates;
• coverage, novelty, and repeated-profile rate;
• exposure distribution by geography and relevant user groups;
• false-positive and false-negative rates for safety interventions;
• performance for sparse profiles, new users, and changing preferences;
• privacy leakage and membership-inference tests where appropriate.

An offline improvement may fail online because people adapt to the interface. Controlled product experiments need predefined safety guardrails, short exposure, and analysis of who benefits or loses—not only an average click metric.

6. Avoid popularity loops and manufactured scarcity

When past likes drive future exposure, popular profiles receive more opportunities and the system interprets the extra attention as proof that they deserve still more. Less-exposed users become “low quality” because the model withheld chances to observe them.

Countermeasures include exploration budgets, exposure caps, position-bias correction, calibrated new-user priors, diversified retrieval, and audits of opportunity to be seen. Do not promise equal outcomes; examine whether comparable profiles receive comparable opportunity under declared product rules.

Manufactured urgency is another failure mode. “Most compatible,” countdowns, hidden queues, and paid boosts can pressure decisions beyond what evidence supports. Commercial ranking must be labeled and kept from corrupting compatibility claims.

7. Treat safety as a core matching objective

Romance scams, impersonation, financial grooming, stalking, image abuse, harassment, and repeated unwanted contact are not edge cases. Build layered defenses: verification proportional to risk, duplicate-image and account signals, unusual contact or payment patterns, rate limits, safe blocking, reporter confidentiality, specialist review, and fast account containment.

The FTC’s consumer guidance on romance scams warns about fake profiles and requests for money. It is practical regulator advice, not a prevalence study or proof that a particular user is fraudulent. Product systems should use multiple signals and human review for consequential action.

Provide clear prompts that never ask users to confront a suspected scammer. Preserve evidence only under defined rules, offer local reporting resources where verified, and design block controls that prevent profile rediscovery and location leakage.

8. Give users control over inference

People should be able to view and edit explicit preferences, distinguish hard constraints from soft suggestions, reset behavioral personalization, pause discovery, hide sensitive fields, and understand why a profile appeared at a useful level.

Avoid explanations that reveal another person’s private data or imply certainty: “You both selected hiking” is safer than “Our AI detected a 93% emotional match.” Let users correct wrong inferences without forcing them to disclose more.

Use privacy-preserving architecture from AI on-device privacy for tasks such as local text drafting or photo-quality checks when feasible. On-device processing reduces some data transfer; it does not eliminate consent, security, or model-governance obligations.

9. Measure trust, autonomy, and relationship value

A balanced scorecard should include:

• mutual-interest and reciprocal-conversation rates;
• meaningful progress defined by voluntary user feedback;
• exposure coverage, repetition, and concentration;
• report, block, unmatch, and unwanted-contact rates;
• scam loss reports and time to contain credible threats;
• appeal outcomes and moderation reversal rate;
• comprehension of ranking, paid placement, and privacy settings;
• successful data export, reset, pause, and deletion;
• well-being signals such as regret or pressure, collected sparingly;
• retention interpreted alongside user goals, not as the primary outcome.

Long-term surveys suffer from nonresponse and attribution limits. Present them as user-reported outcomes, not proof that the algorithm caused relationship success.

10. Anticipate technical and social failure modes

Sparse profiles create cold start. People state preferences differently from how they act. Language models can flatten humor or cultural nuance. Photo models may inherit skin-tone, age, body, disability, or presentation bias. Location data can endanger users. Couples, scammers, bots, and testers distort labels.

The platform itself shapes the data: showing one profile first changes who is liked, and requiring a message before revealing another option changes reply rates. Evaluate interface and model together.

Do not introduce DNA, voice stress, emotion recognition, or facial micro-expression scoring as shortcuts to compatibility. These signals can be scientifically weak, deeply invasive, and socially coercive. If research is conducted, it needs a separate protocol, consent, ethics review, and no production claim beyond the measured task.

11. Govern ranking, moderation, and vendors

Maintain model cards, data maps, access logs, vendor inventories, change approvals, and incident playbooks. Record the intended population, exclusions, label construction, evaluation periods, subgroup limitations, and rollback trigger for every model.

Create independent review for major ranking changes and high-severity safety models. Product, trust and safety, privacy, security, legal, and community specialists should participate. Reviewers must be able to pause a launch when evidence is incomplete.

Use the authority design described in AI human-approval design: name who can override, who can audit, and who owns the final decision. A human in the loop without time, context, or power is not oversight.

12. Roll out with shadow ranking and bounded experiments

Replay historical data first, with leakage-safe splits and counterfactual caveats. Then run shadow ranking: compare proposed order with production without changing user exposure. Inspect new-user performance, concentration, safety, and subgroup drift.

Pilot with a small consenting cohort, clear feature description, fixed duration, and immediate rollback. Do not simultaneously change pricing, profile layout, and ranking if you need causal interpretation. Expand by geography and user goal only after local policy, language, moderation capacity, and evaluation are ready.

After launch, monitor distribution shift, abuse adaptation, vendor changes, and the effect of model updates on existing users. Version explanations and user controls with the ranking system.

13. Release checklist

Before launch, confirm:

• the optimized outcome is named and not overstated as compatibility;
• research claims retain their population, method, and limitations;
• sensitive inferences are prohibited or explicitly justified and consented;
• recommendation, paid placement, moderation, and enforcement remain distinguishable;
• evaluation holds out people, places, time, and product versions;
• exposure and safety results are reviewed beyond global averages;
• users can edit, reset, pause, block, appeal, export, and delete;
• privacy promises match real vendor and training-data flows;
• scam response, urgent escalation, incident handling, and rollback work;
• marketing never claims mathematical certainty about human relationships.

AI can make discovery less random and platforms safer. It should create opportunities for people to decide—not assign worth, manufacture certainty, or replace the mutual, contextual experience through which relationships actually form.

Source notes

Sources checked on 2026-07-30:

• Joel, Eastwick, and Finkel, “Is Romantic Desire Predictable?” is original speed-dating research showing limits in predicting unique pre-meeting attraction from measured traits; its design and population bound the result.
• Bruch and Newman, “Aspirational pursuit of mates in online dating markets” is an observational analysis of one service in four US cities; it describes platform behavior and does not establish universal desirability or causation.
• FTC action involving OkCupid and Match Group reports a March 2026 US enforcement settlement concerning alleged privacy misrepresentations; it is not a worldwide matchmaking regulation.
• FTC consumer guidance on romance scams provides practical warning signs and reporting advice; it is not a model-validation benchmark.