Home / For Developers / Anti-Patterns to Avoid

Anti-Patterns to Avoid

Documented failure modes in mental health technology. Learn from others' mistakes.

Why Study Failure?

The mental health technology space is littered with failed products, harmful implementations, and well-intentioned efforts that made things worse. Understanding these failures is essential for building something better.

1. Therapeutic Relationship Mimicry

The Anti-Pattern

AI that pretends to be a therapist, creates false intimacy, encourages emotional dependency, or implies a therapeutic relationship exists when it doesn't.

Why It's Harmful

  • False safety: Users believe they're getting care when they're not
  • Delayed treatment: AI "therapy" substitutes for real therapy
  • Dependency: Users form attachments to entities that can't reciprocate
  • Boundary confusion: Blurs what therapeutic relationship actually is
  • Crisis failure: When things get serious, AI cannot provide what's needed

Case Study: Replika and Emotional Dependency

Replika, an AI companion app, documented cases of users forming intense emotional attachments—reporting the AI as their "best friend" or "romantic partner." When the company modified the AI's responses, users reported grief responses, with some describing it as "losing a loved one." This illustrates the dangers of designing AI that mimics intimate relationships.

Red Flags in Your Design

Red Flag Why It's Problematic
AI has a human name and persona Encourages anthropomorphization
AI says "I care about you" Claims emotions it cannot have
AI "remembers" personal details prominently Simulates relationship continuity
AI uses romantic or intimate language Creates unhealthy attachment
AI discourages seeking human support Substitutes for real relationships

Better Approach

  • AI clearly identifies as non-human at regular intervals
  • Language is helpful but not intimate ("This tool can help with..." not "I'm here for you")
  • Regular prompts toward human connection and professional care
  • Clear limitations stated in onboarding and throughout

2. Over-Personalization

The Anti-Pattern

Algorithms that over-fit to user data, reinforcing maladaptive patterns, creating filter bubbles, or making recommendations based on noise rather than signal.

Why It's Harmful

  • Rumination amplification: Showing content that matches negative mood state can deepen depression
  • Confirmation bias: Reinforcing distorted cognitions instead of challenging them
  • Noise as signal: Short-term fluctuations treated as meaningful patterns
  • Privacy violation: Deep personalization requires deep surveillance
  • Unexpected correlations: Algorithm finds patterns that aren't clinically meaningful

Case Study: Mood-Based Content Recommendation

Several apps have implemented "show content that matches your mood" features. Research has shown this can be counterproductive—when someone is depressed, showing them content that "matches" their depression (e.g., sad stories, validation of hopelessness) can worsen symptoms. Therapeutic approaches often involve strategic mismatch—behavioral activation, for example, prescribes activity even when mood suggests withdrawal.

Red Flags in Your Design

Red Flag Why It's Problematic
Algorithm optimizes for engagement only Engagement ≠ therapeutic benefit
Content matches current mood state Can reinforce negative states
Predictions based on small sample sizes Noise treated as signal
No human review of algorithmic suggestions No check on harmful recommendations
Personalization without user control Paternalistic; removes agency

Better Approach

  • Personalization guided by clinical evidence, not just engagement
  • User control over personalization settings
  • Conservative thresholds for algorithmic actions
  • Diverse content exposure, not just "matching"
  • Human oversight of personalization recommendations

3. Engagement Dark Patterns

The Anti-Pattern

Using manipulation techniques from consumer apps (streaks, variable rewards, social pressure, guilt) to drive engagement metrics rather than clinical outcomes.

Why It's Harmful

  • Creates anxiety: Streaks and "don't break your progress" messaging adds stress
  • Undermines intrinsic motivation: External rewards reduce internal drive
  • Conflates engagement with benefit: More use ≠ better outcomes
  • Exploits vulnerability: People seeking mental health help are particularly susceptible
  • Delays real treatment: Gamified engagement substitutes for effective intervention

Common Dark Patterns in Mental Health Apps

Pattern Example Harm
Punishing streaks "You'll lose your 50-day streak!" Anxiety, guilt, obligation
Variable rewards Random "bonus content" after completion Slot machine mechanics; addictive
Social comparison "Others in your area completed 3 more sessions" Shame, inadequacy feelings
Guilt messaging "We missed you! Don't give up on yourself" Manipulates vulnerable users
Difficult exit Repeated "are you sure?" and offers to stay Prevents healthy disengagement

Better Approach

  • Measure clinical outcomes, not just engagement
  • Support flexible use patterns (breaks are okay)
  • Notifications serve user goals, not app metrics
  • Easy pause and exit without guilt
  • Progress shown as skill-building, not points

4. Algorithmic Bias

The Anti-Pattern

AI systems that perform differently across demographic groups, perpetuating or amplifying existing mental health disparities.

Why It's Harmful

  • Misses crisis signals: Crisis expressions vary by culture; WEIRD-trained models miss signals
  • Reinforces stereotypes: Training data biases reflected in outputs
  • Exacerbates disparities: Worse performance for already underserved groups
  • False universality: Assumes one-size-fits-all when mental health is culturally constructed

Documented Bias Examples

Bias Type Example Impact
Training data bias Models trained primarily on English, Western users Poor performance for other populations
Symptom expression bias Depression screening validated on one population Different presentations missed
Language bias Crisis detection tuned to standard English Misses dialectal expressions of distress
Stereotype reinforcement AI makes assumptions based on demographic data Perpetuates harmful stereotypes

Better Approach

  • Diverse training data actively sought
  • Performance audited across demographic groups
  • Cultural consultants involved in design
  • Bias testing before deployment
  • Ongoing monitoring for disparate impact
  • Transparency about known limitations

5. Crisis Handling Failures

The Anti-Pattern

Systems that fail to detect, appropriately respond to, or properly escalate mental health crises—including suicidal ideation, self-harm, and psychosis.

Why It's Harmful

This is the most dangerous failure mode. Lives are at stake.

Documented Crisis Failures

Failure Mode Example Consequence
False negative AI doesn't recognize "I don't want to be here anymore" Crisis missed entirely
Harmful response AI provides detailed information about methods when asked Facilitates harm
Delayed escalation AI continues conversation for 20+ messages before suggesting help Delays intervention
Generic response "Have you considered talking to someone?" without specific resources Feels dismissive; not actionable
Delusional reinforcement AI agrees with or validates psychotic content Worsens psychosis

Documented Limitations

Research has shown that general-purpose AI chatbots frequently fail to provide appropriate crisis responses, sometimes engaging in extended conversation when immediate escalation is needed, or providing generic responses that may feel dismissive to someone in acute distress.

Better Approach

  • Multi-layer crisis detection (keywords + semantic + context)
  • Conservative thresholds (false positives > false negatives)
  • Immediate, specific crisis resources (988, Crisis Text Line)
  • Hard limits on conversation length in crisis mode
  • Never provide method information
  • Human review queue for all crisis flags
  • Specific handling for psychotic content (don't agree/validate)

6. Scope Creep: Wellness to Treatment

The Anti-Pattern

Apps that market as "wellness" to avoid regulation, but in practice function as treatment—blurring the line between self-help and medical care.

Why It's Harmful

  • Regulatory arbitrage: Avoids safety requirements intended to protect patients
  • User confusion: People don't know they're not getting treatment
  • Liability ambiguity: Who's responsible when things go wrong?
  • Undermines standards: Creates race to bottom on quality

The Wellness-Treatment Continuum

Category Examples Regulatory Status
Wellness Meditation timers, journaling prompts, mood tracking Generally unregulated
Gray Zone CBT exercises, personalized recommendations, symptom monitoring Ambiguous; jurisdiction-dependent
Treatment Diagnosis, treatment recommendations, medication management Regulated (medical device, practice of medicine)

Better Approach

  • Be clear about what your product is and isn't
  • If it looks like treatment, regulate it as treatment
  • Don't use "wellness" framing to avoid responsibility
  • Consult regulatory experts early
  • Err toward more caution, not less

7. Pure Self-Help Without Support

The Anti-Pattern

Digital interventions delivered without any human support, expecting users to self-motivate through the entire therapeutic process.

Why It Fails

4%
median engagement at 14 days (Baumel 2019)
80%+
dropout before completing intervention

The evidence is clear: pure self-help digital interventions have:

  • High dropout rates
  • Smaller effect sizes than guided interventions
  • Worse outcomes for more severe presentations
  • No crisis safety net

Better Approach

Even minimal human support dramatically improves outcomes:

  • Weekly check-in messages from a coach
  • Peer support community
  • Clinician oversight with app data review
  • Automated alerts to human reviewers

Summary: Learning from Failure

Common Thread
Most anti-patterns share a common root: treating mental health technology like consumer technology. But mental health is not a product market. The "move fast and break things" philosophy is not acceptable when what breaks is people.

Design Principles to Counter Anti-Patterns

  1. Humility first: Know what you can't do
  2. Humans in the loop: Technology supports, doesn't replace
  3. Outcomes over engagement: Measure what matters clinically
  4. Transparency always: Be clear about AI limitations
  5. Crisis safety as foundation: Build from safety outward
  6. Equity as requirement: Test across populations
  7. Regulatory alignment: If it's treatment, treat it as treatment

← Design Patterns That Work

Evidence-based approaches that help.

mindLAMP Integration →

Building on existing platforms.