AI Deepfakes and the Ethics of Synthetic Harm in Science
Deepfakes raise significant ethical questions in the realm of science. Much research relies on trust in data and evidence, and even minor manipulations can lead to serious consequences.
How AI Deepfakes Are Created
Deepfakes are generated using generative models; specifically, neural networks trained on extensive datasets of faces, voices, and movements. These models do not simply replicate existing recordings; they learn the patterns that make individuals look, sound, and behave in specific ways, subsequently using that knowledge to synthesize new, convincing content.
Modern deepfake systems typically utilize one of two primary model types: generative adversarial networks (GANs) and diffusion models.
GANs operate in a sort of digital sparring match. One side consists of a generator attempting to create synthetic content that can pass as real, while the other side is a discriminator working to differentiate real from fake. The generator improves by learning to fool the discriminator, leading to increasingly lifelike outputs over time.
In face-swapping systems, for instance, the model learns a shared representation of facial features, allowing it to overlay one person’s expressions and movements onto another while preserving key elements of their identity.
Initially, GAN-based systems displayed noticeable flaws, such as unnatural lighting or flickering eyes. However, advancements in training data, architecture design, and post-processing have significantly enhanced the quality of today’s outputs, making them harder to distinguish from authentic footage.
Diffusion models, on the other hand, function differently. During training, these models learn how images degrade when noise is added and, crucially, how to reverse that process. When generating content, they start with pure noise and gradually refine it until a realistic image or video frame emerges. This results in smoother, more stable visuals compared to earlier GAN-based methods.
Moreover, diffusion models are particularly versatile as they can be guided by identity information, text prompts, or audio cues, thus making the output highly controllable. However, this flexibility presents new challenges for detection tools, many of which were originally designed to identify specific flaws associated with GANs.
Synthetic Harm and Public Trust
The accuracy of deepfakes has led to rising public apprehension. Much concern has centered on the potential for abuse, including non-consensual intimate imagery and political misinformation, which can erode trust in audio-visual evidence. These issues are particularly pressing as studies have shown that such synthetic harm disproportionately affects women and contributes to increasing public skepticism towards legitimate media.
Beyond personal and political ramifications, deepfakes introduce broader epistemic challenges by destabilizing the credibility of recorded evidence. When individuals can no longer ascertain what is real, trust in journalism, democratic institutions, and scientific data begins to erode.
Scientific Integrity and Synthetic Data
One of the most pressing concerns is the potential misuse of generative AI to produce scientific data and medical imagery that appear authentic but are fabricated. A 2025 PNAS article warns that researchers, companies, or regulators may use generative models to fabricate results that seem methodologically sound, potentially misleading the scientific community.
Risks associated with this include:
- Irreproducible results: Fabricated data may mislead others who attempt to build on it, leading to wasted effort and erroneous conclusions.
- False confidence in findings: Charts, figures, and tables may appear methodologically sound but could be entirely synthetic.
- Privacy breaches: Even when datasets are deemed “synthetic,” if the generative models were trained on real human data, it’s possible to reverse-engineer sensitive personal details.
These scenarios blur the boundaries between real and synthetic harm. For instance, a fabricated MRI image might influence diagnosis or treatment decisions, not out of malice, but due to the absence of ethical safeguards, raising urgent questions about professional trust and evidentiary standards in science.
A Case Study: Deepfakes in Forensic Science
Forensic science is particularly vulnerable, relying heavily on the integrity of digital evidence. Practitioners face increasing challenges in verifying the authenticity of video and audio content that may have been synthetically manipulated.
As deepfakes become more realistic, traditional detection methods, such as identifying visual artifacts or irregular speech patterns, are often insufficient. A 2023 review published in the Journal of Imaging highlighted a surge in academic attention to this issue, noting:
- Threat to evidence credibility: Deepfakes threaten to erode the reliability of digital evidence, potentially compromising the fairness of legal proceedings.
- Detection methods: Current forensic techniques, including Convolutional Neural Networks (CNNs) and spatio-temporal modeling, aim to spot subtle inconsistencies in synthetic media, though they vary widely in accuracy and reproducibility.
- Lack of standardization: A significant limitation across reviewed studies was the absence of a unified framework for detecting and reporting deepfakes in forensic contexts.
The potential for psychological harm and legal uncertainty for victims and professionals alike underscores the need for technological innovation and interdisciplinary collaboration across forensic science, computer vision, ethics, and law.
Regulatory and Ethical Responses
Currently, there are few formal rules directly addressing the use of synthetic media in scientific work. While some legal efforts focus on malicious deepfakes in harassment or political misinformation, the scientific landscape remains murky. Critical questions arise: Was the synthetic data used responsibly? Was it clearly labeled? Or was it misrepresented as real?
Journals, funders, and institutions need to step up their efforts. For instance, publishers could require researchers to disclose when generative tools were employed and build screening processes for altered images or figures. Conferences and preprint platforms could implement checklists or clearer ethics policies.
Ethics boards also have a crucial role as AI tools integrate into everyday research. Questions around consent, privacy, and data generation become more complex, particularly if a synthetic medical image resembles a real patient.
Education plays a vital role as well. Many researchers lack training in identifying synthetic manipulation or understanding its implications. Building AI literacy and digital ethics into scientific training can help individuals use these tools thoughtfully and flag potential issues early.
Given the global nature of science, solutions must also be international, fostering collaboration across borders to create consistent, practical standards for handling synthetic content in research.
Conclusion
The ethical risks of synthetic harm in science cannot be ignored. As deepfake technologies advance and become more accessible, they threaten to undermine the foundational trust that science relies upon. Whether these technologies enhance scientific discovery or erode its foundations depends on our deliberate and responsible handling of them.
