Late at night, a chatbot can start contradicting its earlier identity claims.
The answers can stay fluent.
The continuity can still feel unstable.

Example: A user chats for a long time with an assistant.
The assistant shifts how it describes itself.
The user starts doubting whether it is the same entity.

This piece avoids big claims about a “complete artificial self.”
It organizes design and evaluation conditions for perceived “selfhood” in LLMs.
It focuses on what can be tested with protocols and metrics.
It also separates two targets.
One target is human imitation, like tone and reactions.
The other target is semantic consistency, like memory and goals.

TL;DR

• This separates “human imitation” from “self-consistency” as different evaluation targets in long dialogue.
• It matters because drift and contradictions can weaken trust during multi-turn or multi-session use.
• Next, add LongMemEval, ES-MemEval, or BEAM plus persona-drift tests like Persistent Personas.

Current status

A single standard benchmark for long-term “self-consistency” seems unsettled.
This claim needs further verification.
Two lines of work appear to be developing in parallel.

One line focuses on long-term memory and multi-session memory.
Another line focuses on role and persona maintenance.
It also measures drift, including contradictions.

On long-term memory, one cited example is LongMemEval.
This work describes five core abilities for long-term memory evaluation.
These include information extraction and temporal reasoning.
They also include knowledge update and related abilities.

ES-MemEval is described in a long-term emotional-support setting.
It evaluates memory across five categories.
These include information extraction and temporal reasoning.
They also include conflict detection, abstention, and user modeling.
The aim is not only correct recall.
It also checks abstention when no evidence exists in history.

BEAM targets how performance holds up over time.
It is described as 100 conversations and 2,000 validated questions.
It notes conversations generated up to 10M tokens.
The goal is not only long context length.
It probes how updates, forgetting, and distortion accumulate.

On persona and role maintenance, Persistent Personas is notable.
It uses long persona conversations with more than 100 rounds.
It reports persona fidelity decreases as conversations get longer.

Another thread proposes automatic metrics for persona drift.
The snippet describes three automatic metrics.
They are prompt-to-line consistency, line-to-line consistency, and Q&A consistency.

Analysis

In practice, “artificial self” talk can be operationalized.
The experience can be observed through several measurable items.
One item is long-term persistence across turns or sessions.
Another item is stable self-reference across a conversation.
A third item is traceable goals or values over time.
A fourth item is groundedness of memory claims.

Recent evaluation work seems closer to standardizing memory groundedness.
It also seems closer to persistence under long contexts.
This assessment needs further verification.
Self-reference and goals are often measured indirectly.
Persona drift and consistency metrics are used for this role.

Some user research links mind-like cues to trust outcomes.
A Scientific Reports paper from 2024 is cited in the snippet.
It links perceived experience to altruism in its results.
It links perceived agency to trust in its results.
The snippet provides no effect sizes or numerical results.

Service-robot research also links perceived theory of mind to responses.
The snippet does not provide specific metrics or effect sizes.
A direct within-study comparison was not confirmed in the snippets.
The comparison would separate imitation-centered versus internal-state modeling.
This gap needs further verification.

Turning “self-consistency” into metrics can add distortions.
Improving memory QA accuracy can reinforce plausible reminiscence style.
Improving persona scores can label growth as drift.
Emphasizing abstention can lower satisfaction for some users.
Design often involves trade-offs.
These include consistency versus adaptability.
They also include accuracy versus usefulness.
They also include memory versus privacy.

Practical application

A “build an artificial self” goal can blur system requirements.
It can help to split requirements into two layers.

First is the memory and grounding layer.
Define what to remember and what to ground on.
Define failure behavior, including abstention without grounding.

Second is the identity and goal layer.
Turn persona statements into testable sentences.
Include red lines and persistent goals as testable sentences.

Add imitation details later, like tone and reactions.
Early demos can look plausible if imitation comes first.
Long-run behavior can still wobble without grounding and rules.

Checklist for Today:

• Add one long-term evaluation suite, like LongMemEval, ES-MemEval, or BEAM, plus abstention criteria.
• Convert persona requirements into test sentences, then measure drift with Persistent Personas-style long conversations.
• Track prompt-to-line, line-to-line, and Q&A consistency separately, then review failure cases with humans.

FAQ

Q1. To create a “self,” do we have to deceive users by pretending to be human?
A1. This question is related but separate from deception.
Human-like cues can support early affinity.
Long-term immersion can weaken when continuity breaks.
Tests for memory, grounding, and rule maintenance can reduce this risk.

Q2. What do we use to quantify long-term consistency?
A2. A single widely agreed standard seems unclear.
This point needs further verification.
For long-term memory, examples include LongMemEval, ES-MemEval, and BEAM.
For persona drift, examples include Persistent Personas and automatic metrics.
Those metrics include prompt-to-line, line-to-line, and Q&A consistency.

Q3. Won’t adding “abstention” make the product less useful?
A3. That risk can exist in some product contexts.
In long conversations, false certainty can accumulate trust loss.
One approach is to apply abstention only to grounded items.
These include user settings, past agreements, and sensitive information.
Another approach is UX support for alternatives.
Alternatives include follow-up questions, summaries for confirmation, and options.

Conclusion

I

Further Reading

• AI Automation Shocks Jobs, Energy Costs, Transfer Feasibility
• Bridging the Gap Between AI Performance and Productivity
• Evaluating LLM Operational Reliability Beyond Benchmark Scores
• When Image Preprocessing Breaks Multimodal Geolocation Reliability
• Make AGI Year Predictions Testable With Clear Scoring

References

• LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory - arxiv.org
• ES-MemEval: Benchmarking Conversational Agents on Personalized Long-Term Emotional Support - arxiv.org
• Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in LLMs (BEAM) - arxiv.org
• Persistent Personas? Role-Playing, Instruction Following, and Safety in Extended Interactions - arxiv.org
• Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning - arxiv.org
• Perception of experience influences altruism and perception of agency influences trust in human–machine interactions - nature.com
• Service robots with (perceived) theory of mind: An examination of humans’ reactions - sciencedirect.com