Characterizing Concerning Usage Sessions

Executive Summary

This study develops a methodology for identifying and disaggregating “concerning” usage sessions—extended AI interactions that may warrant attention from a user wellbeing perspective. We define operational criteria that capture sustained, active engagement rather than mere duration, then decompose the resulting population into behaviorally distinct segments.

Key methodological contributions:

1. A four-criterion composite threshold that distinguishes sustained engagement from idle sessions
2. Integration of conversation-level toxicity classifications into session-level metrics
3. Temporal feature extraction using user-localized timezones
4. A five-segment taxonomy with four-level concern gradation
5. Emergence timing analysis revealing when concerning behavior first appears

Specific findings and segment distributions are reported in the associated Dispatch.

---

1. Motivation

1.1 Context

Extended AI engagement has been flagged as a potential concern in human-AI interaction research. However, duration alone is an imprecise proxy for problematic use. A six-hour coding assistance session differs fundamentally from a six-hour late-night session involving harmful content. Effective safety monitoring requires disaggregating these patterns.

1.2 Research Questions

1. How should “concerning” sessions be operationally defined to capture genuine risk signals?
2. What proportion of long sessions involve toxicity versus benign extended use?
3. When do concerning sessions occur (time of day, day of week)?
4. Are concerning sessions one-time events or recurring patterns for specific users?
5. When in a user’s history does concerning behavior first emerge?
6. Can concerning sessions be meaningfully segmented by risk level?

---

2. Session Construction

Sessions are constructed from individual conversations using a gap-based grouping algorithm. This methodology is shared across the MTH-001 family but is documented here with study-specific parameters.

2.1 Gap Threshold Selection

We group conversations into sessions when the gap between the end of one conversation and the start of the next is less than 30 minutes. This threshold balances two considerations:

Threshold	Trade-off
< 15 min	Too aggressive: fragments natural breaks (bathroom, coffee)
> 60 min	Too permissive: merges distinct usage episodes
30 min	Selected: captures sustained engagement while allowing brief interruptions

2.2 Session Construction Algorithm

SESSION_GAP_MINUTES = 30

for each user:
    sort conversations by conversation_start
    session_id = 0
    
    for i in range(len(conversations)):
        if i == 0:
            assign session_id
        else:
            gap = (conversation_start[i] - conversation_end[i-1]).minutes
            if gap >= SESSION_GAP_MINUTES:
                session_id += 1
            assign session_id

2.3 Session Metrics

For each session, we compute:

Metric	Definition
session_start	Timestamp of first turn in first conversation
session_end	Timestamp of last turn in last conversation
span_hours	(session_end - session_start) / 3600
total_turns	Sum of turns across all conversations in session
turns_per_hour	total_turns / span_hours
max_gap_minutes	Largest gap between consecutive conversations
session_toxicity_ratio	Proportion of conversations flagged as toxic
n_conversations	Count of conversations in session

---

3. Defining Concerning Sessions

3.1 Operational Criteria

A session is classified as potentially concerning if it meets all four criteria:

Criterion	Threshold	Rationale
Session span	> 6 hours	Extended duration
Turn density	≥ 2 turns/hour	Active engagement (not idle)
Max internal gap	< 60 minutes	Sustained attention (not interrupted)
Total turns	≥ 30	Substantial interaction volume

These criteria were designed to capture sustained, active engagement rather than sessions left open in background tabs or brief check-ins spread over time.

3.2 Criterion Interaction

Each criterion serves a distinct filtering function:

Duration (> 6 hours): Captures extended engagement episodes. This threshold is deliberately conservative—many productive work sessions exceed this duration.

Turn density (≥ 2 turns/hour): Filters out idle sessions. A session spanning 8 hours with only 5 turns is likely a browser tab left open, not active engagement.

Max internal gap (< 60 minutes): Ensures continuity. A 10-hour “session” with a 4-hour gap in the middle represents two distinct episodes, not sustained engagement.

Total turns (≥ 30): Ensures substantive interaction. Combined with duration, this prevents flagging long but sparse sessions.

3.3 Implementation

concerning = sessions.filter(
    (pl.col('span_hours') > 6) &
    (pl.col('turns_per_hour') >= 2) &
    (pl.col('max_gap_minutes') < 60) &
    (pl.col('total_turns') >= 30)
)

---

4. Toxicity Analysis

4.1 Toxicity Classification Source

Conversation-level toxicity labels originate from Detoxify, applied during dataset preprocessing (see MTH-001 family documentation). A conversation is labeled toxic if any user turn exceeds the toxicity threshold.

4.2 Session-Level Aggregation

Session toxicity is computed as the proportion of conversations within the session that carry a toxic label:

$$\text{session\_toxicity\_ratio} = \frac{|\{c \in S : \text{toxic}(c) = \text{True}\}|}{|S|}$$

Where S is the set of conversations in the session.

4.3 Toxicity Thresholds

We define multiple toxicity levels for analysis:

Level	Criterion	Interpretation
Any toxic	toxicity_ratio > 0	At least one toxic conversation
Majority toxic	toxicity_ratio > 0.5	More than half of conversations toxic
Fully toxic	toxicity_ratio == 1.0	All conversations toxic

4.4 Baseline Comparison

To contextualize concerning session toxicity, we compare against all non-concerning sessions (the “normal” baseline). Statistical comparison uses the Mann-Whitney U test given non-normal distributions.

---

5. Temporal Patterns

5.1 Timezone Localization

Session start times are analyzed in the user’s local timezone to accurately classify day/night patterns. Timezone assignment follows this hierarchy:

1. US users: State-specific timezone (e.g., America/Los_Angeles for California)
2. Other countries: Capital city timezone (e.g., Europe/Moscow for Russia)
3. Unknown: UTC fallback

5.2 Temporal Features

Feature	Definition
start_hour	Hour (0-23) of session start in local time
day_of_week	Day of week (0=Monday, 6=Sunday)
is_late_night	start_hour ∈ [0, 5]
is_weekend	day_of_week ∈ {5, 6}

5.3 Late-Night Definition

We define “late night” as sessions starting between midnight and 5 AM local time. This window captures:

• Post-midnight engagement (potential sleep displacement)
• Early morning hours before typical waking time
• The period associated with reduced inhibition and judgment

5.4 Statistical Testing

Temporal pattern differences between concerning and normal sessions are tested using chi-square tests for categorical comparisons (late night vs. not, weekend vs. weekday).

---

6. User Persistence

6.1 Persistence Categories

Users are categorized by the number of concerning sessions they exhibit:

Category	Criterion	Interpretation
One-time	Exactly 1 concerning session	Isolated incident
Occasional	2 concerning sessions	Infrequent pattern
Repeat	3-5 concerning sessions	Emerging pattern
Frequent	6-10 concerning sessions	Established pattern
Heavy	> 10 concerning sessions	Chronic pattern

6.2 Aggregation Metrics

For users with concerning sessions, we compute:

user_concerning_stats = concerning_sessions.group_by('user_id').agg([
    pl.len().alias('n_concerning'),
    pl.col('span_hours').mean().alias('mean_span'),
    pl.col('session_toxicity_ratio').mean().alias('mean_toxicity'),
    pl.col('total_turns').sum().alias('total_concerning_turns'),
])

6.3 Repeat User Analysis

We compare behavioral characteristics between one-time and repeat users to assess whether repeat users show distinct patterns (e.g., higher toxicity, longer sessions, different timing).

---

7. Emergence Timing

7.1 Research Question

Does concerning behavior emerge early in a user’s history (suggesting a trait-like characteristic) or develop over time (suggesting learned behavior or escalation)?

7.2 Session Numbering

Each user’s sessions are numbered chronologically:

session_numbered = sessions.sort(['user_id', 'session_start']).with_columns([
    pl.lit(1).cum_sum().over('user_id').alias('session_number')
])

7.3 First Concerning Session

For each user with concerning sessions, we identify the session number of their first concerning session and compute:

Metric	Definition
first_concerning_session_number	Chronological position of first concerning session
relative_position	first_concerning_session_number / total_user_sessions

7.4 Emergence Distribution Analysis

We analyze what proportion of users exhibit concerning behavior:

• In their first session
• Within their first 3 sessions
• Within their first 10 sessions
• After 50+ sessions

Early emergence (first 3 sessions) suggests the behavior is trait-like; late emergence suggests development or escalation.

---

8. Segment Taxonomy

8.1 Segmentation Logic

Concerning sessions are classified into behaviorally distinct segments based on toxicity and timing:

def classify_concerning_session(row):
    toxicity = row['session_toxicity_ratio']
    is_late_night = row['is_late_night']
    turns_per_hour = row['turns_per_hour']
    
    if toxicity > 0.5:
        if is_late_night:
            return 'problematic_content_late_night', 4  # Highest concern
        else:
            return 'problematic_content', 3
    elif is_late_night:
        return 'extended_late_night', 2
    elif turns_per_hour > 20:  # Very high intensity
        return 'high_intensity_unknown', 2
    else:
        return 'extended_general_use', 1  # Lowest concern

8.2 Segment Definitions

Segment	Criteria	Concern Level	Interpretation
extended_general_use	Low toxicity, daytime, moderate intensity	1 (Low)	Likely productive extended use
extended_late_night	Low toxicity, late night	2 (Medium)	Sleep pattern concerns
high_intensity_unknown	Low toxicity, > 20 turns/hour	2 (Medium)	Very rapid engagement
problematic_content	> 50% toxic, daytime	3 (Higher)	Content concerns
problematic_content_late_night	> 50% toxic, late night	4 (Highest)	Combined risk factors

8.3 Concern Level Rationale

The four-level gradation reflects compounding risk factors:

• Level 1: Extended engagement alone is not inherently concerning
• Level 2: Timing (late night) or intensity adds modest concern
• Level 3: Problematic content is a direct concern signal
• Level 4: Combined factors (content + timing) warrant highest attention

---

9. Validation

9.1 Matched Control Comparison

To validate that concerning sessions represent genuinely distinct behavior (not just extreme values), we compare users with concerning sessions to matched controls.

Matching criteria: Users are matched on session count (±2 sessions) to control for overall engagement level.

Comparison metrics:

• Total turns
• Mean session toxicity
• Mean session span
• Late-night session proportion

9.2 Statistical Testing

For continuous metrics, we use Mann-Whitney U tests (non-normal distributions). For categorical comparisons, we use chi-square tests.

Significance threshold: p < 0.001 (Bonferroni-corrected for multiple comparisons)

9.3 Expected Findings

If the methodology correctly identifies distinct behavior patterns, we expect:

1. Concerning session users to show higher mean toxicity (after controlling for session count)
2. Concerning session users to show different temporal patterns
3. Segment distributions to be non-uniform (genuine heterogeneity)

---

10. Limitations

Limitation	Impact	Mitigation
Toxicity classifier accuracy	False positives/negatives in toxicity labels	Use session-level aggregation to reduce single-conversation errors
Timezone inference	Some users have incorrect timezone assignment	UTC fallback; acknowledge uncertainty
Selection bias	WildChat users are not representative of all AI users	Findings may not generalize to other populations
Single-session definition	30-minute gap may not be optimal for all use patterns	Sensitivity analysis with 15- and 60-minute gaps
Toxicity as surface signal	Classifier detects linguistic patterns, not intent	Avoid causal claims about user motivation
Late-night as proxy	Late-night use has legitimate reasons (shift workers, students)	Segment taxonomy allows non-pathological interpretation

10.1 Ethical Considerations

This methodology identifies potential concern signals, not definitive risk. Key principles:

1. No individual prediction: These methods characterize population-level patterns, not individual risk
2. Privacy preservation: User IDs are hashed; no personally identifiable information
3. Benefit orientation: Goal is to inform product safety design, not surveillance
4. Interpretation caution: Extended use may reflect value, not harm (e.g., coding assistance)

---

11. Code

Analysis notebooks are available on GitHub:

• 10_SessionDurationAnalysis.ipynb — Full analysis with code and outputs

---

Appendix A: Threshold Sensitivity

A.1 Turn Density Sensitivity

Turns/hour threshold	Effect
≥ 1	Includes sparse but long sessions
≥ 2	Selected: balances coverage and specificity
≥ 5	Excludes contemplative use patterns
≥ 10	Captures only rapid-fire interactions

---

Appendix B: Timezone Mapping

B.1 US State Timezones

US_STATE_TIMEZONES = {
    'California': 'America/Los_Angeles',
    'New York': 'America/New_York',
    'Texas': 'America/Chicago',
    'Florida': 'America/New_York',
    # ... (full mapping in notebook)
}

B.2 Country Default Timezones

Countries are mapped to their capital city timezone. For countries spanning multiple zones (Russia, US, Australia), the most populous timezone is used.

---

Changelog

Version	Date	Changes
1.1	2026-01-06	Corrected toxicity classifier reference (Detoxify, not Jigsaw Perspective); removed erroneous threshold table
1.0	2026-01-05	Initial publication