Immoral Programming: What can be done if malicious actors use language AI to launch ‘deepfake science attacks’?

The problem-solving and imitation capabilities of AI are increasing. In parallel, research addressing ethical AI design has gained momentum internationally. However, from a cybersecurity-oriented perspective in AI safety, it is vital to also analyse and counteract the risks posed by intentional malice. Malicious actors could for instance exploit the attack surface of already deployed AI, poison AI training data, sabotage AI systems at the pre-deployment stage or deliberately design hazardous AI. At a time when topics such as fake news, disinformation, deepfakes and, recently, fake science are affecting online debates in the population at large but also specifically in scientific circles, we thematise the following elephant in the room now and not in hindsight: what can be done if malicious actors use AI for not yet prevalent but technically feasible ‘deepfake science attacks’, i.e. on (applied) science itself? Deepfakes are not restricted to audio and visual phenomena, and deepfake text whose impact could be potentiated with regard to speed, scope, and scale may represent an underestimated avenue for malicious actors. Not only has the imitation capacity of AI improved dramatically, e.g. with the advent of advanced language AI such as GPT-3 (Brown et al., 2020), but generally, present-day AI can already be abused for goals such as (cyber)crime (Kaloudi and Li, 2020) and information warfare (Hartmann and Giles, 2020). Deepfake science attacks on (applied) science and engineering – which belong to the class of what we technically denote as scientific and empirical adversarial (SEA) AI attacks (Aliman and Kester, 2021) – could be instrumental in achieving such aims due to socio-psycho-technological intricacies against which science might not be immune. But if not immunity, could one achieve resilience? This chapter familiarises the reader with a complementary solution to this complex issue: a generic ‘cyborgnetic’ defence (GCD) against SEA AI attacks. As briefly introduced in Chapter 4, the term cyborgnet (which is much more general than and not to be confused with the term ‘cyborg’) stands for a generic, substrate-independent and hybrid functional unit which is instantiated e.g. in couplings of present-day AIs and humans. Amongst many others, GCD uses epistemology, cybersecurity, cybernetics, and creativity research to tailor 10 generic strategies to the concrete exemplary use case of a large language model such as GPT-3. GCD can act as a cognitively diverse transdisciplinary scaffold to defend against SEA AI attacks – albeit with specific caveats.