1. Introduction

Data poisoning is the intentional or reckless alteration, insertion, corruption, or manipulation of data in a dataset, scientific experiment, artificial intelligence (AI) training process, genomic database, clinical trial, or regulatory submission. The objective may be to manipulate analytical outcomes, compromise an AI model, distort scientific conclusions, undermine evidence, or cause an organisation to make an incorrect decision.

Chain of custody is the documented record of how data or evidence was collected, transferred, accessed, stored, analysed, preserved, and presented. Its purpose is to demonstrate that the evidence submitted to a court, regulator, laboratory, or investigator is authentic and has not been improperly altered.

These concepts are closely connected. Data poisoning threatens the integrity of information, while chain-of-custody safeguards help detect, investigate, and prove unauthorised changes.

The legal issues are particularly important in biotechnology, genomics, pharmaceutical research, clinical trials, digital forensics, healthcare AI, and biosurveillance, where unreliable data can affect patient safety, regulatory approvals, intellectual property, criminal investigations, and public health.

Jurisdictional note: The discussion below draws on Indian law, United States decisions, and selected international principles. There is no single universal offence or legal doctrine called “data poisoning”; liability depends on the conduct, intent, affected system, resulting harm, applicable legislation, and jurisdiction.

2. Meaning and principal forms of data poisoning

Accreditation & Quality | SNT Biotech Lab

A. Scientific and research-data poisoning

Manipulating experimental measurements, deleting unfavourable results, changing sample labels, fabricating observations, or altering laboratory records. It may constitute research misconduct, fraud, breach of contract, or a regulatory violation.

Bad data, broken AI: The weak link in today’s AI boom | Technology

B. AI training-data poisoning

Introducing misleading or maliciously selected examples into training datasets so that a model produces inaccurate predictions or behaves incorrectly under particular conditions.

Neue Diagnostik befragt das Gedächtnis des Immunsystems

C. Genomic and biomedical data poisoning

Swapping biological sample identities, manipulating sequence files, altering clinical metadata, or contaminating reference datasets. Such conduct can lead to false diagnoses, invalid research findings, or defective regulatory decisions.

What is digital chain of custody: ensuring legal evidence integrity

D. Digital-evidence tampering

Altering logs, timestamps, device records, electronic documents, or forensic copies to misrepresent what occurred or to undermine the authenticity of evidence in litigation.

An important distinction is that not every data error is poisoning. Accidental contamination, ordinary measurement errors, model bias, poor-quality datasets, and deliberate manipulation are different problems. Intent is relevant to criminal liability and some forms of misconduct, but negligent data handling may independently create civil or regulatory liability.

3. Legal significance of chain of custody

A defensible chain of custody should answer five questions:

Identity: What exactly is the evidence or dataset?

Origin: Who collected it, when, where, and by what method?

Continuity: Who possessed or accessed it at every material stage?

Integrity: What proves that the evidence has not been improperly changed?

Reproducibility: Can an independent expert reconstruct the analysis and verify the result?

For electronic data, this commonly requires a combination of authenticated source records, access logs, cryptographic hashes, reliable timestamps, version histories, and documented analytical procedures.

A cryptographic hash, such as SHA-256, can help show that a file has remained unchanged since a verified reference hash was created. However, it cannot prove that the original file was accurate or unpoisoned. A maliciously altered file can be hashed just as easily as a genuine file. Hashes must therefore be supported by trustworthy acquisition procedures, protected audit logs, and independent verification.

4. Major case laws and their relevance

The following judgments do not all concern AI data poisoning directly. They establish the legal principles most relevant to proving digital manipulation, testing scientific reliability, authenticating electronic records, and maintaining a trustworthy evidentiary chain.

A. Indian case law

1. Anvar P.V. v. P.K. Basheer

Supreme Court of India · (2014) 10 SCC 473

Legal principle: Electronic records offered as secondary evidence were subject to the statutory requirements of Section 65B of the Indian Evidence Act, 1872.

Relevance: A party cannot establish the reliability of a disputed electronic record merely by producing a printout or copied file without complying with the applicable evidentiary requirements. The judgment is foundational for understanding authentication and admissibility of electronic evidence in India.

Current-law qualification: The Bharatiya Sakshya Adhiniyam, 2023 (BSA) replaced the former Evidence Act from 1 July 2024. Section 63 is the principal successor provision governing admissibility of qualifying electronic records. The former law and its precedents remain relevant to interpretation, subject to the current statute and applicable transitional rules.

Indian Kanoon+1

2. Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal

Supreme Court of India · (2020) 7 SCC 1

Legal principle: The Supreme Court reaffirmed the importance of the statutory certificate required under Section 65B(4) for admissibility of electronic records produced as computer output, while clarifying how the certificate may be obtained.

Relevance to data poisoning: In a dispute involving manipulated laboratory logs, sequencing files, emails, or AI audit records, the party relying on electronic copies must satisfy the applicable admissibility rules. Proper certification, however, does not by itself prove that the underlying data is truthful or free from poisoning.

Safeguard: Preserve original records, document extraction procedures, record file hashes, and retain qualified witnesses who can explain how the evidence was generated.

VoxJuris+1

3. Tomaso Bruno v. State of Uttar Pradesh

Supreme Court of India · (2015) 7 SCC 178

Legal principle: The case discussed the significance of available CCTV evidence and the consequences of failing to produce potentially important electronic evidence.

Relevance: If a laboratory, hospital, or technology provider fails to preserve material system logs or recordings that could clarify whether data was altered, the omission may affect the court's assessment of the evidence. An adverse inference is not automatic; it depends on the circumstances and applicable law.

Practical lesson: Establish retention schedules, preserve relevant logs promptly when litigation is anticipated, and document any unavoidable loss of evidence.

Law Maker+1

4. Shafhi Mohammad v. State of Himachal Pradesh

Supreme Court of India · (2018) 2 SCC 801

Legal principle: This decision took a more flexible approach to the certificate requirement for electronic evidence when a party lacked control over the device. That approach was subsequently overruled on this point by the larger-bench decision in Arjun Panditrao.

Relevance: A party should not rely on Shafhi Mohammad as establishing a general exemption from statutory certification. In data-poisoning litigation, evidence should be collected and presented in compliance with the currently applicable evidentiary rules.

Law Maker+1

B. United States case law

5. Daubert v. Merrell Dow Pharmaceuticals, Inc.

United States Supreme Court · 509 U.S. 579 (1993)

Legal principle: Judges must assess whether proposed expert scientific testimony is relevant and rests on a sufficiently reliable scientific foundation.

Relevant considerations include:

Whether the method can be tested.

Whether it has undergone peer review and publication.

Its known or potential error rate.

Whether standards control its operation.

Its acceptance in the relevant scientific community.

Relevance: If an expert relies on an AI model trained on potentially poisoned biomedical data, the court may scrutinise the training data, validation method, reproducibility, error rates, and the basis for the expert's conclusions.

Daubert does not establish a universal rule excluding all evidence associated with data poisoning. It provides an evidentiary framework for evaluating scientific reliability under applicable US federal law.

LII / Legal Information Institute+1

6. General Electric Co. v. Joiner

United States Supreme Court · 522 U.S. 136 (1997)

Legal principle: Appellate courts review a trial court's decision to admit or exclude expert testimony under an abuse-of-discretion standard. The Court also recognised that a court need not accept an expert's conclusions when there is too great an analytical gap between the underlying data and the opinion offered.

Relevance: An expert cannot reliably infer that a biomedical product is safe, effective, or harmful merely by presenting large quantities of data if the methodology connecting that data to the conclusion is unsound.

Practical lesson: Preserve the dataset, analytical code, model version, validation records, and assumptions necessary to test the connection between evidence and conclusion.

LII / Legal Information Institute+1

7. Kumho Tire Co. v. Carmichael

United States Supreme Court · 526 U.S. 137 (1999)

Legal principle: The judicial gatekeeping obligation described in Daubert also applies to technical and other specialised expert testimony, not only conventional scientific testimony.

Relevance: Digital-forensic analysts, cybersecurity experts, genomic-data specialists, and machine-learning engineers may need to explain the technical reliability of their conclusions. Their expertise alone does not eliminate the need for a reliable methodology.

Practical lesson: Document validation procedures, testing results, known limitations, and the steps used to reach the expert opinion.

LII / Legal Information Institute+1

Important limitation: These decisions provide useful evidentiary analogies, but they should not be described as direct judicial rulings on modern AI training-data poisoning unless a particular case actually decided that issue.

5. Legal and regulatory framework in India

Several legal regimes may apply, depending on whether the incident involves electronic evidence, unauthorised computer access, research misconduct, personal data, fraud, or physical harm.

The Information Technology Act is particularly relevant when data poisoning involves unauthorised modification of computer resources. However, not every alteration of data constitutes an offence under Sections 43, 65, or 66. The statutory elements, authorisation, conduct, and required mental state must be examined individually.

Similarly, inaccurate research data does not automatically amount to forgery or cheating. Criminal liability requires proof of the elements of the relevant offence; contractual, professional, scientific-integrity, or regulatory consequences may arise on a different basis.

6. Principal legal risks arising from data poisoning

A. Evidentiary exclusion or reduced weight

If a party cannot authenticate a dataset or establish a sufficiently reliable evidentiary foundation, the court may exclude it where the applicable rules require exclusion, or assign it reduced weight. Admissibility and evidentiary weight are distinct questions.

B. Research misconduct and regulatory consequences

Deliberate fabrication, falsification, or manipulation can trigger institutional investigations, retraction of publications, loss of research funding, employment consequences, or regulatory action. Applicable rules and proof requirements vary by institution and jurisdiction.

C. Patient harm and product liability

Poisoned clinical or genomic data may contribute to incorrect diagnoses, unsafe treatment recommendations, or unreliable safety assessments. Civil liability depends on applicable duties, causation, loss, and available defences.

D. Cybersecurity and privacy violations

An attacker may manipulate personal information, compromise a system, or exploit unauthorised access. Separate data-protection obligations may be triggered by the underlying incident, even where the poisoning itself is not a separately named statutory offence.

E. Contractual and intellectual-property disputes

Altered datasets can undermine representations about research quality, warranties, licence conditions, data-sharing agreements, or regulatory submissions. Remedies depend on the contract, ownership rights, applicable law, and evidence of loss.

7. Essential chain-of-custody safeguards

A robust system should combine technical controls with documented human procedures. No single safeguard is sufficient.

Recommended evidence lifecycle

Collection and identification

Assign a unique evidence identifier. Record the collector, collection time, source device or repository, acquisition method, and relevant sample identifiers. In laboratories, link digital records to physical specimen barcodes and collection records.

Integrity verification at acquisition

Create a cryptographic hash of each relevant file or forensic image. Record the algorithm, hash value, timestamp, operator, and acquisition tool. Where feasible, compare the resulting copy against the original source and document any differences.

Secure storage and access controls

Keep an immutable or tamper-evident audit trail. Restrict access through role-based permissions, multifactor authentication, least-privilege access, and separation of duties. Protect backups and logs from the same administrator or system whose actions they are intended to monitor.

Documented transfers

Record each transfer, including the sender, recipient, date and time, purpose, storage location, integrity checks, and any incident or discrepancy. Use signed transfer records for physical samples and authenticated records for digital transfers.

Reproducible analysis

Preserve the original dataset separately from working copies. Record preprocessing, excluded observations, software and model versions, parameters, dependencies, random seeds where relevant, and any human intervention. Keep the analytical code and sufficient records to reconstruct the result.

Independent validation and closure

Test suspicious records against independent sources, reference samples, original instruments, or trusted backups. Document anomalies, investigation findings, corrective action, and the final disposition of the evidence.

Special safeguards for AI and biomedical datasets

Data provenance: Record where every material training or validation dataset originated, who supplied it, what licence applies, and how it was transformed.

Training-set separation: Maintain independently sourced validation and test sets, with access controls to reduce contamination.

Poisoning detection: Check for anomalous labels, unexpected duplicates, distribution shifts, sample swaps, suspicious clusters, and unexplained changes in model performance.

Model version control: Preserve training-data manifests, model weights or artefacts, code versions, configuration, and evaluation results for each released model.

Human review: Require independent scientific or clinical review before consequential decisions rely on anomalous outputs.

Incident response: Quarantine affected datasets, suspend reliance on potentially compromised models, preserve logs, investigate the attack path, and assess whether notification or reporting obligations apply.

For genomic and clinical research, safeguards should also cover informed consent, specimen identity, laboratory instrument calibration, batch records, chain of sample custody, and controlled access to sensitive information.

8. Hypothetical case study: Poisoned genomic research dataset

Consider a biotechnology company conducting a clinical study using genomic data to develop an AI-assisted diagnostic system.

The company discovers that a group of records was deliberately altered before model training. Some patient labels were switched, and the final model produces unreliable risk predictions.

The legal investigation would proceed along the following lines.

Illustrative investigation workflow

Preserve evidence. Secure original sequencing files, laboratory logs, model artefacts, access logs, and relevant communications. Prevent routine deletion of potentially material records.

Verify integrity. Compare available hashes and historical versions, examine audit logs, and determine when the suspicious changes occurred.

Establish responsibility. Examine access permissions, authenticated activity records, transfer histories, and other corroborating evidence. An account name alone does not prove who performed an action.

Assess scientific impact. Reproduce the training process with verified data and determine whether the suspected poisoning changed the model's performance or clinical conclusions.

Determine legal consequences. Consider research-integrity proceedings, contractual remedies, regulatory reporting, patient notification where required, and civil or criminal action where the relevant legal elements are satisfied.

How the cited cases would assist

Arjun Panditrao would inform the admissibility requirements for relevant electronic records under the former Evidence Act framework and its statutory successor.

Tomaso Bruno would support examination of the consequences of failing to produce material electronic evidence, where the circumstances justify it.

Daubert and Joiner would offer useful US evidentiary principles for testing the reliability of expert conclusions and their connection to the underlying data.

These cases would not, by themselves, establish that a particular employee committed data poisoning. The investigator would still need reliable evidence of alteration, its origin, its significance, and the applicable elements of liability.

9. Evidence-preservation checklist

Use this checklist when preparing an internal investigation, regulatory response, or litigation-readiness review.

Review checklist

0 of 12 completed

Evidence preservation

Identify original datasets, source devices, physical samples, and relevant backups.

Create documented forensic copies and record acquisition details.

Preserve audit logs, timestamps, metadata, and relevant communications.

Integrity and authentication

Record cryptographic hashes and verify them against trusted reference values.

Document each transfer, access event, and change to working copies.

Obtain applicable electronic-evidence certificates and supporting witness statements.

Scientific and technical review

Test for anomalous records, label changes, contamination, and unexplained discrepancies.

Preserve model versions, source code, preprocessing steps, and evaluation results.

Commission independent validation where the evidence or outcome is disputed.

Governance and response

Restrict access and separate investigative duties from routine system administration.

Document findings, corrective actions, and any gaps in the chain of custody.

Assess notification, reporting, contractual, and regulatory obligations.

Reset checklistCopy checklist

10. Conclusion

Data poisoning and chain-of-custody failures present related but distinct legal problems. Data poisoning can compromise the truthfulness and scientific validity of information; chain-of-custody failures can make it harder to establish the authenticity, continuity, and reliability of that information in legal proceedings.

The strongest legal and technical response combines:

Secure acquisition and preservation of original evidence.

Tamper-evident audit trails and independently verified integrity checks.

Reproducible analysis and independent scientific validation.

Compliance with applicable electronic-evidence and sector-specific laws.

Documented investigation, accountability, and proportionate corrective action.

In India, Sections 61–63 of the Bharatiya Sakshya Adhiniyam, 2023 and the relevant provisions of the Information Technology Act, 2000 are particularly important. The Digital Personal Data Protection Act, 2023 also requires attention to its phased commencement: the fact that the Act and Rules have been notified does not mean that every substantive provision was already in force on 4 November 2026.

Indian Kanoon+3

The central legal principle is that preserving a record's integrity is not the same as proving its truth. A defensible chain of custody must establish where the evidence came from and how it was handled, while scientific and forensic validation must separately establish whether the evidence can be trusted.

This is a general legal research overview, not jurisdiction-specific legal advice. Actual liability and admissibility depend on the governing law, the nature of the evidence, and the facts of the particular dispute.