PRISM: Confronting a Future with Conscious Machines
Blog by Will Millership (CEO of PRISM) and Radhika Chadwick (Non-Executive Chair of PRISM)
Last week marked an exciting milestone as we launched PRISM - the Partnership for Research Into Sentient Machines - with a workshop at AI UK, the national showcase of AI hosted by The Alan Turing Institute.
We were overwhelmed by the enthusiasm for our inaugural workshop - The Path to Responsible Conscious AI. Spaces were booked up within 15 minutes of going live and we had a queue down the corridor for drop-ins on the day! The topic of AI consciousness clearly resonated with conference goers, and a concern repeatedly voiced was the urgency of raising these conversations now.
With rapid advancements in AI, the window for ensuring safe and ethical approaches to consciousness may be narrowing, and participants expressed a fear of the potential existential risks to humanity if these issues are not proactively addressed.
In the workshop, we explored some of the key theories of consciousness and how they relate to artificial consciousness. We then discussed the principles adopted by PRISM for responsible research. Finally, we workshopped how to progress today from principles into action to mitigate risks and create a safer future for humanity.
A diverse set of perspectives
We opened the session by finding out what motivated our participants, and we were delighted with the variety of perspectives. We heard from a lawyer, historian, cognitive scientist, statistician, banker, biomedical expert, developer, and data processor. Some of the questions raised were:
What are the geopolitical implications of conscious AI?
Will we need entirely new legal frameworks if machines become conscious?
If AGI is unavoidable, how do we ensure it is ethical and compassionate?
Is conscious AI a potential existential risk? If so, how do we avoid it?
What are the ethical implications of creating potentially conscious video game characters?
How can we “deal with” consciousness without impacting innovation?
Our participants came with varied motivations, united by a deep curiosity and a sense of urgency to address these issues proactively and collaboratively to avoid potential catastrophic outcomes.
Exploring consciousness
In the first two discussions, we explored some of the broad approaches taken to the problem of consciousness, the difficulties of measuring consciousness in living beings, and how (over time) consciousness has been attributed to more and more creatures.
Discussions highlighted that the lack of consensus on the definitions of consciousness could cause difficulty going forward. As one participant (a Geographer by training) put it: “It’s like map making, how can we know where to go, if we cannot agree on where we are.” Other participants pointed to the anthropocentric nature of many attempts to measure consciousness, with humans often assumed to be at the “top” of the consciousness scale, and others pointed out the difficulties of measuring different states of consciousness within individuals.
Among the groups, it was widely agreed that current AI systems are not conscious and that large language models are simply “reflecting the consciousness of the data they are using.” In other words, they are mimicking human behaviour rather than feeling true subjective experiences.
However, this consensus came with a critical warning: as AI becomes more sophisticated, the line between mimicry and genuine consciousness may blur rapidly, raising profound ethical questions sooner than anticipated.
Why does it matter?
While most agree that current AI systems are not conscious, Butlin and Lappas point out that “the features of the brain that are responsible for consciousness, according to some prominent neuroscientific theories, are likely to be reproducible in AI systems.”
Furthermore, Anil Seth (while skeptical of the possibility of machine consciousness) has suggested that “real artificial consciousness is unlikely along current trajectories, but becomes more plausible as AI becomes more brain-like and/or life-like.”
This suggests we may be approaching a critical threshold. Without proactive steps today, we risk finding ourselves unprepared for the complex moral, ethical, and potentially existential dilemmas that true artificial consciousness might introduce.
For our final discussions, we drew on some of the ideas in the paper by Butlin and Lappas on the consequences of over-attributing and under-attributing consciousness. They outline the need to think about the ethical treatment of conscious AI systems as well as the social significance of attributing consciousness to AI systems that aren’t.
While participants acknowledged the need to think about how conscious AI systems will be treated, there was a pressing concern about how humans would be treated in a society with conscious machines.
Questions included:
How do we ensure that machines with consciousness are kind to us and treat us with dignity and respect?
How do we “install” morals and values that are aligned with ours into conscious AIs?
Who is accountable when a human is harmed by a conscious AI, on purpose or by mistake?
Participants emphasised the need to answer these questions of accountability and alignment clearly, stressing that failing to do so could result in harm to individuals and erosion of public trust in AI technologies.
Principles for Responsible AI Consciousness Research
Our final discussions revolved around five principles drawn up by Butlin and Lappas (and adopted by PRISM) to guide any organisation engaged in research that could lead to the creation of conscious machines. They are:
Objectives: Organisations should prioritise research on understanding and assessing AI consciousness with the objectives of (i) preventing the mistreatment and suffering of conscious AI systems and (ii) understanding the benefits and risks associated with consciousness in AI systems with different capacities and functions.
Development: Organisations should pursue the development of conscious AI systems only if (i) doing so will contribute significantly to the objectives stated in principle 1 and (ii) effective mechanisms are employed to minimise the risk of these systems experiencing and causing suffering.
Phased approach: Organisations should pursue a phased development approach, progressing gradually towards systems that are more likely to be conscious or are expected to undergo richer conscious experiences. Throughout this process, organisations should (i) implement strict and transparent risk and safety protocols and (ii) consult with external experts to understand the implications of their progress and decide whether and how to proceed further.
Knowledge sharing: Organisations should have a transparent knowledge sharing protocol that requires them to (i) make information available to the public, the research community and authorities, but only insofar as this is compatible with (ii) preventing irresponsible actors from acquiring information that could enable them to create and deploy conscious AI systems that might be mistreated or cause harm.
Communication: Organisations should refrain from making overconfident or misleading statements regarding their ability to understand and create conscious AI. They should acknowledge the inherent uncertainties in their work, recognise the risk of mistreating AI moral patients, and be aware of the potential impact that communication about AI consciousness can have on public perception and policy making.
If you agree with these principles, you can sign Conscium’s open letter.
There was broad agreement with (and enthusiasm for) the principles. One participant suggested including more about the processes for accountability when things go wrong, and there were discussions around which principles were relevant for different organisations. This led nicely to our final exercise.
Open letter from Conscium.
Principles into Action
In our final session, we assigned each of our six groups a key player in the ecosystem and discussed their responsibilities and obligations. We asked the groups to come up with practical actions these players could take today to avoid potential pitfalls and advance responsible approaches.
Here are some of their thoughts.
UK Government
Participants saw the strength of the UK Government in its “soft power.” They acknowledged that it is not leading the way in development but could position itself as a leader in AI safety, ethics, and governance, “acting as a counterbalance to leading governments that might be neglecting AI ethics.”
Suggested actions:
Establish a Department of AI Ethics within the government.
Build collaborations with the EU, other governments, and non-governmental organisations to strengthen its “soft power.”
Draw up laws for AI safety for product-based companies working with AI.
Found a “Supreme Court for AI” to ensure correct interpretation and application of laws.
Support research into understanding AI consciousness and safety, and encourage link-building between the private sector and academia.
A Global Agency on AI Consciousness
There was a lot of enthusiasm for an international agency that could “get the right players around the table,” lead discussions on the implications of conscious AI, and “create an agreement on actions.”
Concern was raised about ensuring power as an international agency and getting players to agree on guidelines. The NICE Guidance (evidence-based recommendations for the health and social care sector) was suggested as a potential model. The UN Declaration of Human Rights, created by representatives with different legal and cultural backgrounds and widely adopted, was also put forward as an inspiration. The group also suggested that the agency may want to take a philosophical and creative approach to understanding consciousness and had concerns that two agencies may be needed, one to protect the rights of AIs and the other to protect the rights of humans.
Suggested actions:
Work on the definition of terms and standards of AI consciousness.
Create a “Declaration of Rights” for sentient machines as a foundation for legal and political systems.
Build a “register to identify entities” that are conscious or showing signs of consciousness.
Build a register of companies working on AI consciousness or are likely to create it.
Lead informed discussions of AI consciousness with the public and key stakeholders.
Leading Tech Companies
The group representing large tech companies had a pessimistic view of their intentions. They voiced concerns that tech companies are likely to “move fast and break things” in pursuit of their goals - “above all to make money.” They insisted that big tech companies are likely to “push hard against the limits to innovation,” but if they could be seen as "responsible and caring,” it could be good for business.
Suggested actions:
Lead the way in investing in AI alignment research.
Conduct rigorous simulations before releasing AI systems with potentially conscious properties.
Provide clear and honest communications on AI capabilities not over or under attributing consciousness.
Set up internal AI ethics boards.
Agree to third-party audits to ensure compliance with safety standards.
Leading Countries
The group representing the leading countries (USA and China) had three recommendations:
Regulate: Regulate AI systems but make the regulations open for public participation, guided by a range of stakeholders.
Safety governance: “Doomsday planning” for the worst-case scenario of an AI consciousness arms race. Remain ready for all possibilities, especially the worst.
Protect data: The USA and China are the main players when it comes to data infrastructures (data centres, internet cables, etc). Therefore, they should focus on protecting citizens’ data.
Academic Institutes
Participants saw the role of academic institutions as leading research that could help decision-makers make informed decisions, particularly around moral and ethical questions of conscious AI.
Suggested actions:
Conduct impact studies to understand the potential consequences of conscious AI, both positive and negative.
Work on the definitions of consciousness and on different tests that can be used to evaluate artificial consciousness.
Use “jailbreak” techniques to help test potential guardrails and identify common problematic patterns.
The Media
It was acknowledged that consciousness in AI is a delicate subject and that the media has a large part to play. Participants believed that “media hype and sensationalism” could lead people to believe AIs are conscious when they are not and to misguided efforts to protect machines. They called for “unbiased and balanced reporting” to ensure informed discussions.
Suggested actions:
Engage in “AI literacy” programmes to ensure journalists are up to date on the definitions and evaluations of consciousness.
Set up protocols for handling misinformation and disinformation. These might include:
Reporting mechanisms for the public to use.
Identification and traceability of information and sources.
Ways to independently verify information and a standardisation of verification.
Establish robust consequences for publishing incorrect information.
During discussions, many stakeholders highlighted the critical nature of these actions, noting that delay could result in catastrophic consequences, including irreversible harm to society, loss of public trust, and even existential risks.
Reflections
The breadth of expertise and enthusiasm from our participants underscored just how multifaceted this issue is—spanning ethics, law, governance, philosophy, and technology.
While there was consensus that today’s AI systems are not conscious, discussions made it clear that we must proactively prepare for the possibility that future AI systems may develop consciousness. The principles for responsible AI consciousness research offer a strong foundation, but as we saw in our final exercise, translating principles into action requires cooperation across various sectors.
Participants strongly emphasised that delay or neglect in addressing these questions now could lead to severe repercussions and that we need a variety of points; it cannot be left to the companies and governments leading AI development.
Next Steps
As PRISM moves forward, we aim to continue fostering these critical conversations and to move from principles into concrete actions. The challenge ahead is not just to develop safeguards for potentially conscious AI but also to navigate the societal and ethical implications in a way that prioritises innovation and responsibility.
If we can approach this with openness, interdisciplinary collaboration, and a commitment to ethical foresight, we stand a much better chance of shaping a future where AI - whether conscious or not - works in harmony with human values.
This is the beginning for PRISM, and we invite you to be part of this journey! After strong demand, we will be setting up a community where our followers can continue the discussions.
Sign up for our newsletter to be notified when it goes live and to find out about our other planned activities.
