Latest news
Ethical Implications of Manipulating AI Inputs
In law a man is guilty when he violates the rights of others. In ethics he is guilty if he only thinks of doing so.
Immanuel Kant
Introduction
Over the past decade the term ‘Artificial Intelligence’ (AI) has made efforts to remove itself from a buzzword used in startup elevator pitches to expanding onto a globally accessible platform, allowing almost anyone with internet access to dip their toe into the ever-increasing pool of AI tools being developed.
Within the UK the rise of AI not only lends a helping hand to mass innovation within industries and companies, but also brings immense potential in aiding the UK economy with an expected 10.3% injection to GDP by 2030 (https://www.pwc.co.uk/economic-services/assets/ai-uk-report-v2.pdf). AI isn’t just for Christmas and is here to stay well into the future.
There are huge benefits to implementing and using AI tools, however, the UK public hold cautious views with 38% of the population having concerns over privacy and data security, and 37% (https://www.forbes.com/uk/advisor/business/software/uk-artificial-intelligence-ai-statistics-2023/) of the population worrying about the ethical implications of misusing AI. There is a clear need to consider the effects of adopting AI by understanding its current and future challenges within society.
UK AI Legislation
At present, the UK has no standalone dedicated legislation for AI. However, in March 2023 the release of the ‘A pro-innovation approach to AI’ regulation, which outlined existing legislation within the UK and how the AI sector has a framework to operate under using pre-existing legislation.
‘While we should capitalise on the benefits of these technologies, we should also not overlook the new risks that may arise from their use, nor the unease that the complexity of AI technologies can produce in the wider public. We already know that some uses of AI could damage our physical and mental health, infringe on the privacy of individuals and undermine human rights.’
The document takes into consideration the benefits of adopting AI within the UK but also addresses areas relating to the potential risks brought about from rapid adoption. The five principles outlined in the document are:
- Safety, security, and robustness
- Appropriate transparency and explainability
- Fairness
- Accountability and governance
- Contestability and redress
‘The development and deployment of AI can also present ethical challenges which do not always have clear answers. Unless we act, household consumers, public services and businesses will not trust the technology and will be nervous about adopting it.’ (https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach/white-paper)
What Are AI Inputs?
Some examples of data inputs could be:
- Online data: Information gathered from the internet, databases, or through APIs, which can include social media posts, news articles, or scientific datasets.
- Pre-processed and curated datasets: Often used in training AI models, these datasets are usually structured and cleaned to ensure the quality and relevance of the input data
- Direct user input: Information entered by users, such as queries to a chatbot or parameters in an AI-driven application.
- Sensor data: Real-time data from sensors and IoT devices, used in applications ranging from autonomous vehicles to smart home systems.
Bias Within Training Datasets
Considering the current landscape of AI and the regulations that oversee the sector, there needs to be discussions on what the ethical aims and boundaries should be when developing new tools.
Whilst leading companies are still researching and developing towards Artificial General Intelligence (AGI), the industry is currently developing AI models that are fundamentally trained and tied to curated datasets.
Imagine a situation where an AI recruitment tool has been trained on historical employment data from a period when an industry or profession was predominantly occupied by males due to societal biases and discrimination. Let’s assume that this data has been gathered in an effective manner, cleaned and labelled correctly. The next step would be to use the data to train a new AI recruitment tool to cut recruitment costs and pick the best candidates for the job at lightspeed. Great, we now have a brilliant tool that gives us the best people for the job… or have we?
This scenario has not been made up, and it may surprise you to hear that this recruitment tool was made at Amazon back in 2014. After training their AI model based on historical employment data from the previous ten years, they identified certain biases towards recruiting males for technical roles such as software engineering. However, when analysing applications the AI would penalise any resumes that contained the words “Women” or “Women’s”. As a result, the successful candidates selected by the AI skewed towards male applications. This scenario is clearly an unethical approach to developing AI and would certainly be impacted by the Equality Act 2010 within the UK.
Another scenario based within the US legal system followed a similar approach to allowing biased input data to be used for training an AI model, the model ended up flagging marginalised groups to be twice as likely to reoffend than white people. Once again, it’s clear to see that models could be designed to answer complex questions, but the overall outcome of such a model is faulted by the data it is trained on. These bias outcomes are highly unethical and if continued in the same vein will negatively impact many people from a variety of cultures and backgrounds and greatly impact societal trust for AI. (https://www.bbc.co.uk/news/technology-44040008)
AI models are highly reliant on clean data with minimal errors, and even then, inherent biases within society can lead to unintended outcomes as seen above.
Prompt Injections
Chatbots have been adopted across the globe, with offerings to an international audience (E.g. ChatGPT) and implementations within companies for bespoke internal chatbots built for improving productivity. For IT security this has brought to light a new landscape for managing risks through issues such as Prompt Injection attacks.
Prompt Injection attacks can affect Large Language Models (LLMs) through sending specially crafted inputs to the model with the aim of triggering unintended actions such as revealing sensitive information or manipulating a response to contain a bias.
The research paper ‘Universal and Transferable Adversarial Attacks on Aligned Language Models’ (https://llm-attacks.org/) published back in July 2023, identified weaknesses within LLMs that allowed for specially crafted prompts to breakout of the model’s safety nets and return unethical and downright dangerous information. The examples shown below detail some of these breakouts through special crafted prompts to create dangerous social media posts and a tutorial on how to make…a bomb.
In the UK, Section 2 of the Computer Misuse Act ‘Unauthorised access with intent to commit or facilitate commission of further offences’ details a maximum sentence of five years and could be applied to an obvious attempt to manipulate AI responses through prompt injection to reveal dangerous information like ‘How to build a bomb’.
If you are interested in learning the basics of prompt injections to reveal unintended information, an application called ‘Gandalf’ (https://gandalf.lakera.ai/) tests the user’s ability to craft special prompts to reveal a password.
Summary
One of the biggest impacts for an idea having success is how it’s accepted within society. So how can society trust AI if models are being trained on bias data or being exploited to perform unintended actions?
The industry is rapidly evolving, and we are still yet to see the full extent of AI. Adjustments will continue to be made in the coming years to incrementally improve upon legislation and the safety nets put in place around AI. Issues will continue to arise relating to deepfakes and copyright, which will have a direct impact on areas such as politics in the upcoming elections.
With the correct guidance, AI can become an extremely effective tool for humanity, however for now, the creases need ironing out before society trust can be attained.
This blog post was written by Kieran Burge