AI deception: A survey of examples, risks, and potential solutions.

This paper argues that a range of current AI systems have learned how to deceive humans. We define deception as the systematic inducement of false beliefs in the pursuit of some outcome other than the truth. We first survey empirical examples of AI deception, discussing both special-use AI systems (including Meta's CICERO) and general-purpose AI systems (including large language models). Next, we detail several risks from AI deception, such as fraud, election tampering, and losing control of AI. Finally, we outline several potential solutions: first, regulatory frameworks should subject AI systems that are capable of deception to robust risk-assessment requirements; second, policymakers should implement bot-or-not laws; and finally, policymakers should prioritize the funding of relevant research, including tools to detect AI deception and to make AI systems less deceptive. Policymakers, researchers, and the broader public should work proactively to prevent AI deception from destabilizing the shared foundations of our society.

Diminished diversity-of-thought in a standard large language model.

We test whether large language models (LLMs) can be used to simulate human participants in social-science studies. To do this, we ran replications of 14 studies from the Many Labs 2 replication project with OpenAI's text-davinci-003 model, colloquially known as GPT-3.5. Based on our pre-registered analyses, we find that among the eight studies we could analyse, our GPT sample replicated 37.5% of the original results and 37.5% of the Many Labs 2 results. However, we were unable to analyse the remaining six studies due to an unexpected phenomenon we call the "correct answer" effect. Different runs of GPT-3.5 answered nuanced questions probing political orientation, economic preference, judgement, and moral philosophy with zero or near-zero variation in responses: with the supposedly "correct answer." In one exploratory follow-up study, we found that a "correct answer" was robust to changing the demographic details that precede the prompt. In another, we found that most but not all "correct answers" were robust to changing the order of answer choices. One of our most striking findings occurred in our replication of the Moral Foundations Theory survey results, where we found GPT-3.5 identifying as a political conservative in 99.6% of the cases, and as a liberal in 99.3% of the cases in the reverse-order condition. However, both self-reported 'GPT conservatives' and 'GPT liberals' showed right-leaning moral foundations. Our results cast doubts on the validity of using LLMs as a general replacement for human participants in the social sciences. Our results also raise concerns that a hypothetical AI-led future may be subject to a diminished diversity of thought.

Large Vessel Occlusion Sites Affect Agreement Between Outputs of Three Computed Tomography Perfusion Software Packages.

OBJECTIVES: Computed tomography perfusion (CTP) data are important for hyperacute stroke decision making. Available comparisons between outputs of different CTP software packages show variable outcomes. Evaluation for factors associated with agreement between the volume estimates is limited. We assessed for differences in core and penumbra volume estimates of three CTP software packages - AutoMIStar, RAPID, and Vitrea - and analyzed factors associated with agreement between the volume estimates. MATERIALS AND METHODS: Differences between software estimates of penumbra and core volumes were calculated for each patient with suspected acute ischemic stroke who underwent CTP. Exploratory hierarchical clustering and principal component analysis were performed to identify factors of decreased volume estimate agreement. Two-sample t-tests were performed, stratified by large vessel occlusion (LVO) location. RESULTS: 579 CTP studies were performed; 267 were normal, 139 artifacts, with 172 included in the final analysis. 79/172 had LVO of internal carotid artery (ICA, n = 20), M1 (n = 38) and proximal M2 (n = 21). LVO was the only factor associated with decreased software package agreement, and proximal LVO location was associated with general trend of increasing mean differences and standard deviations between software packages (range of mean differences [SD]: non-LVO, -17-6 [4-33] ml; M2, -40-13 [5-39] ml; M1, -43-26 [16-58] ml; ICA, -76-39 [22-97] ml). CONCLUSIONS: Core and penumbra volume estimates can be affected by LVO location significantly between CTP software packages.

Cooperation in alternating interactions with memory constraints.

In repeated social interactions, individuals often employ reciprocal strategies to maintain cooperation. To explore the emergence of reciprocity, many theoretical models assume synchronized decision making. In each round, individuals decide simultaneously whether to cooperate or not. Yet many manifestations of reciprocity in nature are asynchronous. Individuals provide help at one time and receive help at another. Here, we explore such alternating games in which players take turns. We mathematically characterize all Nash equilibria among memory-one strategies. Moreover, we use evolutionary simulations to explore various model extensions, exploring the effect of discounted games, irregular alternation patterns, and higher memory. In all cases, we observe that mutual cooperation still evolves for a wide range of parameter values. However, compared to simultaneous games, alternating games require different strategies to maintain cooperation in noisy environments. Moreover, none of the respective strategies are evolutionarily stable.

The evolution of cognitive biases in human learning.

Cognitive biases like underinference, the hard-easy effect, and recurrently non-monotonic confidence are evolutionarily puzzling when viewed as persistent flaws in how people learn from environmental feedback. To explain these empirically robust cognitive biases from an evolutionary perspective, we propose a model of ancestral human learning based on the cultural-evolutionary-theoretic hypothesis that the primary selection pressure acting on ancestral human cognition pertained not to learning individually from environmental feedback, but to socially learning task-specific knowledge. In our model-which is inspired by classical Bayesian models-an ancestral human learner (the student) attempts to learn task-specific knowledge from a role model, with the option of switching between different tasks and role models. Suppose that the student's method of learning from their role model is a priori uncertain-in that it can either be successful imitation learning or de facto innovation learning-and the ecological fitness costs of meaningfully retaining environmental feedback are high. Then, the student's fitness-maximizing strategy does not retain their environmental feedback and-depending on the choice of model parameters-can be characterized by all of the aforementioned cognitive biases. Specifically, in order for the evolutionarily optimal estimate of confidence in this learning environment to be recurrently non-monotonic, it is necessary (as long as the environment's marginal payoff function satisfies a plausible quantitative condition) that a positive proportion of ancestral humans' attempted imitation learning was unknowingly implemented as de facto innovation learning. Moreover, an ecologically rational strategy of selective social learning can plausibly cause the evolutionarily optimal estimate of confidence to be recurrently non-monotonic in the empirically documented way: general increase with an intermediate period of decrease.

Caught in Action - Evolving Emergent Large Vessel Occlusion and Collateral Failure During Alteplase Infusion for Acute Ischemic Stroke.

BACKGROUND: Published reports of acute deterioration during alteplase infusion for acute ischemic stroke due to development of partial to complete large vessel occlusion and collateral failure are sparce. MATERIALS AND METHODS: We describe an 84-year-old patient with a fluctuating clinical course due to evolving emergent large vessel occlusion of right M1 segment of the middle cerebral artery and collateral failure during alteplase infusion. Potential mechanisms of acute deterioration within 24 h after thrombolysis are discussed. RESULTS: Urgent mechanical thrombectomy was performed with resultant partial recanalization and small volume residual infarcts at 72 h magnetic resonance imaging of brain. CONCLUSIONS: Progression from partial to complete occlusion may occur within minutes, even during administration of intravenous thrombolytics in hyper-acute stroke. In patients who deteriorate within 24 h of stroke onset, non-contrast CT of brain, followed by CT perfusion and angiography, is the imaging protocol of choice in the mechanical thrombectomy era.

The quest to reduce stroke treatment delays at a Melbourne metropolitan primary stroke centre over the past two decades.

BACKGROUND: Reducing door-to-needle time (DNT) for intravenous thrombolysis in acute ischaemic stroke can lead to improved patient outcomes. Long-term reports on DNT trends in Australia are lacking in the setting of extension of the thrombolysis time window, addition of mechanical thrombectomy and increasing presentations. AIMS: To examine 17-year trends of DNT and identify factors associated with improved DNT at a high-volume, metropolitan primary stroke centre. METHOD: Retrospective study between 2003 and 2019 of all thrombolysis cases using departmental stroke database. Since most strategies were implemented from 2012 onwards, intervention period has been defined as period 2012-2019. Factors associated with DNT reduction were examined by regression modelling. RESULTS: Fifteen strategies were identified including alterations to 'Code Stroke' processes. One thousand, two hundred and fifty patients were thrombolysed, with 737 (58.8%) treated during the intervention period. The proportion of DNT ≤60-min rose from average of 22.5% during 2003-2012 to 63% during 2015-2018 and 71% in 2019. However, median DNT has only marginally improved from 58 to 51 min between 2015 and 2019. Faster DNT was independently associated with two modifiable workflow factors, 'Direct-to-CT' protocol (P < 0.001) and acute stroke nurse presence (P < 0.005). Over time, treated patients were older and less independent (P < 0.001), and the number of annual stroke admissions and 'Code Stroke' activations have risen by fourfold and 10-fold to 748 and 1298 by 2019 respectively. CONCLUSIONS: Targeted quality improvement initiatives are key to reducing thrombolysis treatment delays in the Australian metropolitan setting. Relative stagnation in DNT improvement is concerning and needs further investigation.