However, despite these opportunities, it is important to ensure that the design and deployment of AI systems remain fair and equitable. As AI becomes increasingly enmeshed within our personal and professional lives, risks to individuals and society at large have emerged. One notable risk is unfair treatment by systems managed by algorithms with disproportional biases that may lead to a reduced ability to participate in society meaningfully. The term 'bias' can mean different things in different domains. Statistical bias ¹ when an operation is disproportionately weighted to favour some outcome. Social bias happens when such operations relate to people, which may lead to unfair decisions.

Reducing bias is challenging due to the complexity of data and models, as well as potential differing views on whether something is socially biased, even if it may be statistically accurate. For example, men as a group are physically stronger than women as a group. However, we typically consider gender to be a protected characteristic, which is not permitted to unduly influence decisions in hiring, etc. This means that a system may be technically correct yet still problematic in the eyes of the law.

Even data put through a high-pass filter to obfuscate inaccurate machine perceptions to the degree that a human could never recognise it may still contain signatures that machine learning can recognise ²

Bias can sneak in from a number of sources, for example:

1. The reproduction of human labelling or selection biases, such as an algorithm trained upon human appraisals of resumes, may replicate the same biased patterns ³

2. Bias due to errors in datasets, for example, geolocation data that wrongly states that a house is inside a lake and therefore considered unviable for insurance  ⁴

3. Bias due to a lack of sampling data, for example, an algorithm that is trained with a set of examples over-representative of one ethnicity or gender ⁵

4. Bias due to overfitting, whereby a model is trained too much on training data to the degree that it maps poorly onto real-world examples.

5. Bias due to adversarial error, where a model may fail to recognise something accurately or may misinterpret one thing for another ⁶ Models can be reverse-engineered to uncover such exploits.

We can take several steps to reduce the risk of bias within algorithmic systems.

1. Select data which appears to be minimally influenced by human perception or prejudice. This is challenging as data generally needs to be labelled and annotated to be interpreted by machine intelligence.

2. Make datasets more inclusive. Ensure that data is gathered from as broad a sampling as possible, and indeed solicit fewer common examples to ensure that the data is more representative of a global population and global environments.

3. Ensure data accuracy and integrity as far as possible. Perform tests to ensure 'sanity checks' upon data to search for signatures of error and to attempt to locate lacunae (missing data) and either repair it or ideally set it aside. This kind of work is a core duty of data science, and much of these rather dull efforts are performed by legions of workers in less-developed nations for very small sums, with uncertain credentials or quality control ⁷

4. Rigorously test models against real-world examples. Often, a portion of training data is set aside to validate that the model is learning correctly. However, much like a battle plan only lasts until the first engagement with the enemy, lab results are not trustworthy. Systems must be tested live in as broad a range of environments and demographics as possible to be validated as truly accurate and effective.

5. Harden systems against attack and exploitation. Resources should be ring-fenced to provide bounties for red teams to attempt to disrupt the algorithmic system. This can help uncover issues long before they may occur 'in the wild' where real people may be affected. Machine learning systems are increasingly enmeshed with our personal and professional lives. We interact with algorithms a hundred times a day, usually without even realising. It's crucial that such technologies are not applied to exclude anyone or allowed to unfairly misinterpret people's behaviour or preferences.

It's crucial that we embed transparency within algorithmic systems so that we can understand what processes are being performed, in what manner, for what purposes, and to whose benefit. This can help provide insights regarding biases within such systems.

AI has tremendous potential within our society, but there are also risks of it turning into a prejudiced petty tyrant. More governmental, academic, and business resources must be devoted to ensuring that we integrate AI safely and securely into our global society.

Regulating AI fairness

Many factors influence the probability of regulatory effects upon catastrophic risks to fairness and economic franchise, with several trade-offs.

Risk reduction factors: Standard setting: Regulations can set the bar for greater responsibility and accountability, and even standards can become soft law if incorporated into government tenders or embedded with established practices and industry professional credentials. Improved standards and professionalism within industries can lead to improved governance and record-keeping.

Public safety and liability: The availability of insurance, security red teams, and crisis management facilities will tend to limit less catastrophic risks and may provide early warnings of imminent greater disaster. Compounding iterations: The more developments in AI safety are made, generally the greater likelihood of developing the knowledge infrastructure necessary to mitigate catastrophic risk. The more that basic research into AI safety is undertaken and funded, with career opportunities in a newly established formal research discipline, the greater likelihood of discovering advances that pave the way for eventual reduced catastrophic risks.

Commercial opportunities: A marketable safety improvement presents a competitive advantage, even if it may not be very meaningful. Establishing benchmarks for safety which can be applied within comparison and promotional materials can provide incentives for innovation and improved standards.

Risk increasing factors

Obfuscation: Regulations may drive research underground where it is harder to monitor or to ‘flag of convenience’ jurisdictions with lax restrictions by embedding dangerous technologies within apparently benign cover operations (multipurpose technologies). Or by obfuscating the externalised effects of a system, such as in the vehicle emissions scandal  ⁸

Arms race: Recent advances in machine learning, such as multimodal abstractions models (aka Transformers, Large Language Models, Foundation Models) such as GPT-3 and DALL-E, illustrate that dumping computing resources (and the funds for them) in colossal models seems to be a worthy investment. So far, there is no apparent limit or diminishing return on model size, and so now state and non-state actors are scrambling to produce the largest models feasible to access thousands of new capabilities never before possible. An arms race is afoot. Such arms races can lead to a rapid and unexpected take-off in terms of AI capability, and the rush can blindside people to risks, especially when the loss of a race can mean an existential threat to a nation or organisation.

Perverse incentives: Incentives can be powerful forces within organisations, and financialisation, moral panic, or fear of political danger may cause irrational or incorrigible behaviour of personnel within organisations.

Postmodern warfare: Inexpensive drones and other AI-enabled technologies have tremendous disruptive promise within the realm of warfare, especially given their asynchronous nature. Control of drone swarms must be performed using AI technologies, and this may encourage the entire theatre of war to be increasingly delegating to AI, perhaps including the interpretation of rules of engagement and grand strategy ⁹

Cyber warfare: Hacking is increasingly being augmented with machine intelligence, through GAN-enabled password crackers and advanced social engineering tools ¹⁰ This is equally the case in the realm of defence, where only machine intelligence may provide the swift execution required to defend systems from attack. A lack of international cyber war regulation, and poor international policing of organised cybercrimes, may increase the risk of catastrophic risks to societal systems.

Zersetzung: The human mind is becoming a new theatre of war through personalised generative propaganda, which may even extend to gaslighting attacks on targeted individuals, significantly leading to the destabilisation of societies. Such technologies are also plausibly deniable, being difficult to prove who may be responsible.

Inflexibility: After WW1, the German Military was not allowed to develop their artillery material and so developed powerful rocket technologies instead, as these were not subject to regulation. Similarly, inflexible rules may permit exploitable loopholes in AI. They may also not be sufficiently adaptive to implement new technologies and even improved industry standards.

Another example is how the Titanic was permitted to sail without enough lifeboats for everyone due to a primitive Board of Trade algorithm. It calculated lifeboat required based upon tonnage and cubic feet of accommodations, which became outdated due to scaling factors as ship sizes increased. It was also due to a limited lookup table in the regulations that stopped at 10,000 tons and was not updated.

The inverse could also occur. A rule that ‘any model with a parameter size greater than n must…’ could become meaningless if models become much more efficient or if parameters cease to be an applicable measure of model power.

Inflexibility can also manifest where a solution to a problem is found, which then becomes broadly accepted as best practice, anchoring against better solutions being innovated or adopted ¹¹

Limitation of problem spaces: It may be taboo to allow machine intelligence to work on sensitive issues or to be exposed to controversial (if potentially accurate) datasets. This may limit the ability of AI to make sense of complex issues and thereby hinder solutions to crises.

Wilful ignorance: AI may be prevented from perceiving ‘biases’ that are actually uncomfortable truths due to political taboos. For example, it might be prevented from perceiving women as being physically less strong than men as a group, and such a blind spot could produce strange behaviour, potentially leading to runaway effects.

Conclusions

Greater transparency and accountability should be major factors in reducing catastrophic risk. All things being equal, it should be easier to know about the ethical risks of systems, as well as who is culpable for any externalised effects such as disproportional bias.

On balance, I would expect regulation to be generally beneficial to AI ethics, as long as it is not too inflexible, restrictive or overly politicised. It is very important that technology regulation NEVER becomes a polarising issue. Broad, bi-partisan support must be developed if it is to be successful. Otherwise, a substantial proportion of the population will ignore it, whilst the other greater part applies it as a cudgel to harm people by wilfully taking their behaviour out of its proper context to unfairly label them as antisocial.