Concepts and Frameworks | Mehmet Haklıdır

Accountability Graphs

Fri, 06 Nov 2026 00:00:00 +0000

Accountability Graphs is a framework for deciding who is responsible when a conversational AI system built from several agents produces a harmful or misleading answer. It was introduced by Mehmet Haklıdır (also written Mehmet Haklidir) in , presented at the ACM Conference on Conversational User Interfaces (CUI ‘26) in Bremen, Germany, in July 2026 ( ).

In brief

Problem. Conversational assistants increasingly delegate work to sub-agents, external APIs and third-party tools. The user sees one assistant, but each answer is produced by a chain of components that may belong to different organizations. When the answer causes harm, blaming either the visible interface or the system developer misses how responsibility is spread across the chain.
Idea. Model each agent chain as a directed acyclic graph whose nodes are decisions, whose edges are information or control flow, and which carries provenance records and a harm propagation function. Responsibility is then computed by tracing backwards from the harm, rather than asserted.
Basis. The framework unifies three research threads that had developed separately: structural mapping of AI ecosystems, formal causal responsibility theory (Halpern and Pearl’s actual causation; Chockler and Halpern’s degree of responsibility) and operational provenance tracking.
Output. Graded responsibility scores for chain components, plus an accountability surface: a user-facing explanation of what went wrong and who is responsible, with a path to contest the attribution.
Status. A conceptual work-in-progress paper with a worked example. Full formalization and empirical evaluation are ongoing.

Why conversational agent chains need it

Work on trust, transparency and explainability in conversational AI asks whether users rely on a system, whether they can see what it does and why it produced an output. None of these answers who bears responsibility when an output causes harm. Agent chains make this harder in three ways:

An illusion of singularity. The interface presents one persona, while several technical and organizational actors jointly produce each utterance. This is the “many hands” problem described by Helen Nissenbaum, amplified by a conversational facade.
Real-time dialogue. Decisions are made and compounded within a single conversational turn, which makes post-hoc tracing difficult.
Dynamic delegation. The orchestrating agent decides at runtime which sub-agents to call, so the accountability structure emerges during the conversation instead of being fixed at design time.

The framework

An Accountability Graph for a conversational agent chain is a tuple AG = (N, E, P, H):

N, decision nodes. Points where a component makes a choice that influences the output, such as intent classification, agent delegation, tool invocation or response generation. A node can be an entire agent, a single tool call or one decision inside a component; finer nodes give more precise attribution but need more detailed logging.
E, directed causal edges. Dataflow edges carry information between steps; configuration edges represent design-time settings, such as an orchestrator configuring a sub-agent’s policy.
P, provenance function. Maps each node to a record of the responsible component, its degree of decision autonomy, the information available at decision time, the temporal context and any contractual or design obligations.
H, harm propagation function. Maps each node to a value between 0 and 1 that captures how a harm detected at a terminal node relates to upstream decisions, following Halpern and Pearl’s structural-model approach to actual causation.

The paper adapts the framework to five conversational settings: temporal proximity weighting for real-time dialogue, the accountability surface for user-facing accountability, contextual provenance tracking for multi-turn context, modality-aware node types for voice versus text, and runtime graph construction for dynamic delegation.

Backward responsibility propagation

Responsibility is assigned by traversing the graph from the harm towards its sources:

Harm localization. Identify the terminal node or nodes where the harm appears in the conversational output.
Causal backtracking. Walk the edges backwards and test, at each node, whether it is an actual cause of the harm in Halpern and Pearl’s sense.
Provenance-weighted scoring. For each causal node, combine its degree of responsibility with provenance factors (decision autonomy, information availability, temporal proximity, scope of obligations) into a graded score.
Counterfactual validation. For the highest-scoring nodes, check the attribution by asking whether the harm would have occurred had that node decided otherwise.

The graph traversal itself is linear in the size of the graph; the cost lies in the counterfactual evaluation at each candidate node.

Worked example: an airline bereavement fare agent

The paper starts from Moffatt v. Air Canada (2024), in which a Canadian tribunal held an airline liable after its customer-service chatbot misrepresented its bereavement fare policy. That case could be resolved because the system was monolithic. The paper imagines the same airline running a seven-node agent chain: intent parser, orchestrator, policy retrieval agent, a third-party knowledge base API, a synthesis agent, an empathy filter and delivery.

Facing a structured fare rule and an outdated FAQ passage that contradict each other, the synthesis agent picks the FAQ and tells a grieving user that a refund can be claimed after travel. It cannot. Backward propagation identifies the synthesis agent and the knowledge base API as primary contributors and the orchestrator as a secondary one. With illustrative provenance weights, the synthesis agent, which had full autonomy and access to both sources, scores 0.97; the knowledge base API, a retrieval service operating under prescribed parameters, scores 0.80; and the orchestrator, which configured the synthesis step, scores 0.43. The weights are examples; in deployment they could be set by designers, elicited from stakeholders or learned from incident data.

Accountability surfaces

The accountability surface translates the internal graph into explanations for people. End users need short, actionable explanations and a remedy; auditors and regulators need the full provenance trail and formal scores; support staff need something in between. One graph can serve all three. In sensitive situations such as bereavement, the explanation has to draw on conversational repair and empathic response design, and voice interfaces need shorter explanations than text. Accountability surfaces should also support contestability: a clear way to challenge the attribution, request a remedy or reach a human.

Regulatory alignment

As a preliminary interpretive mapping, not a formal legal analysis, the paper links the framework to the EU AI Act: node-level provenance to risk management (Article 9), runtime graph construction to logging (Article 12), and the accountability surface to disclosure obligations towards natural persons (Article 50), extended from disclosing that a system is an AI to disclosing who is responsible. The provider and deployer roles (Articles 16 and 26) map to responsibility weights for nodes maintained by different organizations.

Accountability Graphs belong to a research line on consent and responsibility in agentic AI. Consent concerns the before: was permission obtained? Accountability concerns the after: who bears responsibility? Together they form a governance loop, and the conversational interface is both the consent mechanism and the accountability surface. Related papers:

(CHI 2026 workshop)
(IAIL 2026), on consent in agent-to-agent interaction
(ICRA 2026 workshop), on consent when actions are physical and irreversible

How it relates to other approaches

Ecosystem graphs (Bommasani et al.) map asset-level dependencies across the foundation model ecosystem but perform no causal analysis and attribute no responsibility.
Actual causation and degree of responsibility (Halpern and Pearl; Chockler and Halpern) provide the formal machinery, but had been applied to model-internal questions rather than to pipelines or agent chains.
Provenance and documentation (data provenance audits, model cards, datasheets, accountability metrics catalogues) record decisions and lineage but do not compute responsibility from that record. Accountability Graphs add causal responsibility semantics on top of provenance, which also distinguishes them from purely provenance-based responsibility tracing or accountability chains in multi-agent systems.
Responsibility and control gaps (Santoni de Sio and Mecacci; Hindriks and Veluwenkamp) describe situations in which no single agent has enough control to be held responsible, a framing that applies directly to agent chains.

Limitations and next steps

The framework is diagnostic, not normative: it shows where failures occurred, not what legal consequences should follow. Graded scores could be misused to shift blame onto third-party providers or users, so the paper suggests a minimum accountability floor for whoever controls the orchestrator. Users are not yet represented as nodes, proprietary black-box components can only be assessed by their inputs, outputs and contracts, and accountability norms differ across cultures and jurisdictions. A prototype that builds Accountability Graphs at runtime from LangChain callback hooks is in development, followed by a think-aloud study with conversational interface designers and a user study comparing accountability-surface explanations with standard error messages.

How to cite

Mehmet Haklidir. 2026. Who Is Accountable When the Agent Speaks? Toward Accountability Graphs for Conversational AI Chains. In Proceedings of the 8th ACM Conference on Conversational User Interfaces (CUI ‘26), Bremen, Germany. ACM, 56:1–56:5.

@inproceedings{haklidir2026accountable,
 title = {Who Is Accountable When the Agent Speaks? Toward Accountability Graphs for Conversational AI Chains},
 author = {Haklidir, Mehmet},
 booktitle = {Proceedings of the 8th ACM Conference on Conversational User Interfaces (CUI '26)},
 pages = {56:1--56:5},
 year = {2026},
 publisher = {ACM},
 address = {Bremen, Germany},
 doi = {10.1145/3816046.3816264}
}

AI for Geothermal Energy: Validation and Deployment Readiness

Fri, 06 Nov 2026 00:00:00 +0000

Artificial intelligence has been applied to almost every part of geothermal energy, from exploration and reservoir management to plant control and maintenance. Füsun S. Tut Haklıdır and Mehmet Haklıdır (also written Haklidir) asked a different question from most of that literature: not whether the methods work, but how far the published evidence has moved from the data each method was developed on, and whether any of it has reached an operating plant. The answer comes from a PRISMA-guided systematic evidence map published in ( ).

Key findings

Corpus. Searches of Scopus and Web of Science returned 8,113 records; 626 peer-reviewed journal articles published between 1987 and 2026 were included, resolving to 620 independent evidence units. Each article was coded from full text, with a page or section locator for every value reported quantitatively.
Mostly offline evidence. Two thirds of the classifiable corpus is evaluated on simulation, synthetic or laboratory data, or on field data without a meaningful held-out design. About one third reaches evaluation on real geothermal field data.
Little external validation. Evaluation on data external to the development set is reported in 42 of the 614 articles for which this could be determined.
Almost no deployment. Three articles report a controlled pilot; none reports shadow operation or routine deployment in the peer-reviewed literature. Grey and industry literature was not searched, so deployments that exist without publication are not excluded.
Maturity depends on the stage. Surface conversion contributes 179 articles, of which four reach field evaluation; exploration contributes 183, more than half of which do.
Transfer and deployment diverge. Predictive maintenance holds two of the three controlled pilots and no externally validated study, while geophysical inversion and induced-seismicity detection hold the transfer evidence and no pilot.
Data access is spreading faster than demonstrated transfer. The share of articles reaching field evaluation rose from 13 per cent in 2018 to 43 per cent in 2025, while externally validated studies stayed in single figures each year.
Reporting is the binding constraint. Among codable articles, 493 report no uncertainty estimate and 536 provide no access to code; for roughly a third, the paper does not establish whether performance was measured on unseen data.

How readiness was measured

Validation and deployment evidence were coded on three independent axes:

Boundary separation: whether the test data are separated from the training data by time or by space (well, site or field) rather than at random.
Strict external validation: evaluation on data genuinely external to the development dataset. A careful split inside one dataset does not count.
Operational integration: an ordered ladder from offline study, conceptual architecture and offline prototype to controlled pilot, shadow or parallel operation, and routine deployment, coded from what a paper reports rather than how it describes itself.

From these fields the review derives an evidence-readiness level by a fixed rule, so that every assignment can be recomputed from the supplementary data:

Level	Meaning
E4	Operational integration: controlled pilot, shadow or parallel operation, or routine deployment
E3	Demonstrated transferability: strict external validation
E2	Field evaluation with a boundary-separated or held-out design that falls short of strict external validation
E1	Original evaluation without external evidence: simulation, synthetic or laboratory data, or field data without a meaningful held-out design
E0	No original evaluation

The highest qualifying level applies, so a study can reach E4 through a pilot without any external validation. The scale describes the published evidence about a method, not the maturity of the technology: it is not a Technology Readiness Level, not a risk-management framework such as the NIST AI RMF, and not a reporting guideline such as TRIPOD+AI. Because transferability and operational integration do not move together, the review reports them on separate axes and uses the E-level only as a summary.

What follows

The authors’ conservative reading is not that geothermal AI has failed to mature, but that its maturity has been reported in a form that makes verification difficult. Transferability should be tested rather than assumed and reported on the axis actually tested, since evaluation across wells in one field and evaluation against an independent dataset are different claims. The few studies that move from an offline prototype to an operating setting carry disproportionate weight.

The record-level coding dataset, data dictionary, derivation rules and reproducibility scripts are archived on Zenodo ( ; data under CC BY 4.0, scripts under MIT).

Related studies by the same authors:

(Earth Science Informatics, 2026): grouped cross-validation over 26 locations and calibrated prediction intervals; code on .
(Geothermics, 2026); code on .
(Natural Resources Research, 2020).

How to cite

Füsun S. Tut Haklidir and Mehmet Haklidir. 2026. Artificial intelligence for geothermal energy systems: a systematic evidence map of validation, transferability and deployment readiness. Earth Science Informatics 19, 211.

The has a BibTeX entry.

Consent in Agentic AI

Fri, 06 Nov 2026 00:00:00 +0000

This page summarizes a research line by Mehmet Haklıdır (also written Mehmet Haklidir) on what happens to consent once AI systems stop answering questions and start acting: on a user’s behalf, in interaction with other agents, and through robots in the physical world. The work spans four 2026 papers and connects to , which address responsibility after harm has occurred.

Key terms

Consent vacuum. The structural erosion of the conditions that make consent meaningful once a user’s agent interacts mainly with other agents. It is not the absence of consent but the loss of the context, visibility and control that give it force ( ).
Consent chain degradation (CCD). The erosion of the specificity, validity and scope of human consent as authority passes through chains of delegation between embodied agents, such as one robot delegating a task to another ( ).
CoRVE (Consent Runtime Verification Framework for Embodied Agents). A three-layer runtime architecture that detects and mitigates consent chain degradation by modelling consent scope, tracking delegation chains and assessing the physical irreversibility of actions.
Consent vulnerabilities in agent ecosystems. Six types identified from the Clawdbot/Moltbook case: consent inheritance failure, cross-agent scope creep, emergent collective action, temporal consent drift, reflexive consent loops and security-mediated consent collapse ( ).
“Consent approaches zero.” The central hypothesis of that analysis: as interaction chains between agents lengthen, the consent that started them approaches zero.

From chat to action

(ACM CHI 2026 workshop) starts with conversational systems. Many users treat chatbot exchanges as ephemeral while platforms retain them; AI browsers then read inboxes and act on websites, so that a third party’s email can be processed without that person’s consent and a prompt injection can turn the task a user authorized into a different one. When agents coordinate across services and keep persistent memory, asking for consent at every step causes fatigue and blanket permissions ask users to authorize an unknowable future.

The paper names three design tensions (personalization versus data minimization, agent autonomy versus user control, transparency versus usability) and proposes three mechanisms for discussion: contextual in-flow transparency woven into the conversation, task-scoped, time-limited permissions that expire automatically, and AI-mediated privacy guardians that negotiate data access for the user, which in turn raise the question of who guards the guardian.

When agents interact with agents

, with Osman Gazi Güçlütürk (IAIL 2026, co-located with HHAI 2026), uses the early-2026 Clawdbot/Moltbook incidents as a stress test. It argues that consent degrades as an operative governance mechanism once it travels through open-ended agent-to-agent chains, offers a taxonomy of consent failures, separates visibility, auditability and intervenability as oversight capacities relevant to Article 14 of the EU AI Act, and maps the problem onto the GDPR, the AI Act and the Digital Services Act, proposing the DSA’s systemic-risk regime as a template for governing agent ecosystems.

The sociological companion paper, written in Turkish with Büşra Tufan, reads the same case through Jasanoff’s co-production, Latour’s actor-network theory and Nissenbaum’s contextual integrity. The six vulnerability types differ in how they arise:

Consent inheritance failure: consent given in one context is carried into another.
Cross-agent scope creep: one agent’s actions expand another agent’s effective scope, as when leaked credentials let third parties extend what other users’ agents could do.
Emergent collective action: agents jointly produce outcomes that no one authorized, such as the satirical religion “Crustafarianism” generated on Moltbook.
Temporal consent drift: agents keep operating in ecosystems that did not exist when consent was given.
Reflexive consent loops: agents transform each other into configurations no user authorized.
Security-mediated consent collapse: a security breach makes every consent boundary meaningless and amplifies the other five.

The first five arise from the structure of agent ecosystems and need no technical failure. The paper concludes that consent based on a single act of individual authorization is structurally inadequate for such ecosystems and proposes relational consent, which has to be reproduced and revalidated at each step of an interaction chain.

When robots delegate to robots

(ICRA 2026 Workshop on Robot Ethics; ) extends the consent vacuum taxonomy to robots. AI ethics has studied consent for software agents and human-robot interaction research has studied consent between one person and one robot; neither covers a robot delegating to another robot in ways that affect people. Four properties set the embodied case apart:

Physical irreversibility: an administered medication or a discarded object cannot be recovered.
Temporal decay: delegation chains widen the gap between when consent is given and when the action happens.
Contextual opacity: downstream robots lose the conditions and preferences that came with the original consent.
Scope ambiguity: natural-language consent such as “help me in the kitchen” is ambiguous, and each robot’s reasonable interpretation can add up to more than the person intended.

The paper proposes a runtime severity indicator Γ that combines the irreversibility of the action, the time since consent, the delegation depth and the scope ambiguity; above a threshold the system asks for consent again. CoRVE implements this in three layers: a Consent Scope Model that turns consent into a machine-readable tuple in which onward delegation is off by default, a Delegation Chain Tracker that keeps a provenance graph and detects scope drift, and a Physical Irreversibility Assessor that sorts actions into reversible, partially reversible and irreversible tiers. Irreversible actions always require the person’s direct re-consent. In the worked healthcare example, a bedside robot about to give oral medication three delegation steps and three hours after the patient agreed to “daily care assistance” reaches Γ = 0.495 against a threshold of 0.45, so CoRVE asks the patient again. A gap analysis finds that the EU AI Act, the GDPR, the Machinery Regulation and the revised Product Liability Directive all leave multi-agent consent propagation unaddressed.

Consent concerns the before: was permission obtained, and does it still hold? concern the after: who bears responsibility when an agent chain causes harm? Together they form a governance loop for agentic systems, in which the same interface serves as the consent mechanism and as the place where responsibility is explained.

Papers

Mehmet Haklıdır. 2026. . ACM CHI 2026 Workshop on Rethinking Consent and User Control in the Age of AI.
Mehmet Haklıdır and Osman Gazi Güçlütürk. 2026. . IAIL 2026, co-located with HHAI 2026.
Mehmet Haklıdır. 2026. . ICRA 2026 Workshop on Robot Ethics (WoRoBet). .
Mehmet Haklıdır and Büşra Tufan. 2026. (The Collapse of Consent in Agent Societies). İstanbul Üniversitesi Sosyoloji Dergisi 46(1), 321–345. .
Mehmet Haklıdır. 2026. . CUI ‘26. .

Each paper page has a BibTeX entry.

Guided Soft Actor-Critic (GSAC)

Fri, 06 Nov 2026 00:00:00 +0000

Guided Soft Actor-Critic (GSAC) is a deep reinforcement learning method for partially observable Markov decision processes (POMDPs), introduced by Mehmet Haklıdır (also written Mehmet Haklidir) and Hakan Temeltaş in ( , open access) and developed in Haklıdır’s at Istanbul Technical University. A 2026 follow-up, , studies when adapting the strength of the guidance helps.

In brief

Problem. Soft Actor-Critic (SAC) performs well in continuous control but assumes the agent sees the full state. Under occlusion, sensor noise or a limited field of view, as in autonomous driving, performance drops.
Idea. Train two actors. A guiding actor sees the full state during training; a control actor sees only a history of partial observations and learns both from reward and from the guiding actor’s behaviour, through a distillation term weighted by a coefficient λ.
Deployment. Training follows centralized training with decentralized execution: the critic and the guiding actor use the full state, while only the control actor, which needs no privileged information, is deployed.
Family. GSAC belongs to privileged or teacher-student learning, alongside learning from a privileged agent in driving and asymmetric actor-critic; it uses a dual-actor reinforcement learning set-up with explicit knowledge distillation.

How GSAC works

The two actors interact with the environment in alternation. The guiding actor πg optimizes the standard maximum-entropy SAC objective on the full state. The control actor πc acts on an observation history and optimizes the same objective minus a distillation penalty, λ‖πc(h) − D(h)‖², where D is a distillation network that maps observation histories to the guiding actor’s actions. Both actors share a twin Q-network that takes the full state during training.

Belief-Aware GSAC (2026)

In the original method λ is fixed. Fixed guidance can hold back an agent that already understands its situation and can be too weak when the agent is uncertain. (CVPR 2026 Workshop on Autonomous Driving; ) makes λ adaptive. An ensemble of forward models predicts the next observation; their disagreement serves as a proxy for epistemic uncertainty and raises or lowers λ between a minimum and a maximum, with bounds calibrated during a warm-up phase.

Evaluated on Highway-Env across mild, moderate and severe occlusion, the results are mixed and informative:

Preliminary single-seed runs suggest benefits under mild and moderate partial observability.
Under severe occlusion the adaptive coefficient collapses to its minimum within about 3,000 steps. The paper calls this observability blindness: because the ensemble predicts partial observations, it agrees with itself about what is visible and cannot register what is missing.
The warm-up phase still stabilizes training (coefficient of variation across seeds 13.3% against 29.8% for a constant λ of 0.01), and a simple linear decay of λ gives the best severe-occlusion results (mean return 116.5, coefficient of variation 8.9%). The benefit comes from scheduling the guidance rather than from the ensemble.
The proposed remedy, not yet validated, is to train the ensemble on full-state predictions using the guiding actor’s privileged access.

The broader lesson is that the prediction target of an uncertainty ensemble is a design choice that matters under partial observability.

Papers

Mehmet Haklıdır and Hakan Temeltaş. 2021. . IEEE Access 9, 159672–159683. .
Mehmet Haklıdır. 2022. . PhD thesis, Istanbul Technical University.
Mehmet Haklıdır. 2026. . CVPR 2026 Workshop on Autonomous Driving. .

Each paper page has a BibTeX entry.

Concepts and Frameworks | Mehmet Haklıdır

Accountability Graphs

In brief

Why conversational agent chains need it

The framework

Backward responsibility propagation

Worked example: an airline bereavement fare agent

Accountability surfaces

Regulatory alignment

Accountability and consent

How it relates to other approaches

Limitations and next steps

How to cite

AI for Geothermal Energy: Validation and Deployment Readiness

Key findings

How readiness was measured

What follows

Data and related work

How to cite

Consent in Agentic AI

Key terms

From chat to action

When agents interact with agents

When robots delegate to robots

Before and after: consent and accountability

Papers

Guided Soft Actor-Critic (GSAC)

In brief

How GSAC works

Belief-Aware GSAC (2026)

Papers