LoginAn Intelligent Approach to AI
For better and worse, artificial intelligence could revolutionise our working lives. But away from the hype and scare-mongering, how can Trading Standards make use of it?
Cenred Elworthy, CTSI Lead Officer for Digital, AI and Emerging Markets
Posted 26 May 2025 | JoTS Online
Content Tags: Analysis|Technology|National
Discussions about the use of artificial intelligence (AI) tools by government and law enforcement – including Trading Standards – are often rife with heightened expectations. In part these stem from excessive investment-led hype, and in part from politically motivated claims about the potential for massive cost efficiencies that could arise if only AI were more widely introduced. This is in addition to valid concerns about data protection and the accuracy of AI-produced materials. In this short article I intend to share some recent insights into the most advanced AI models, consider possible legal implications arising from AI agency and then, hopefully, land softly back down to earth with practical tips on how to make the most of AI tools.
Thinking allowed
In his seminal 1974 paper, ‘What Is It Like to Be a Bat?’, the philosopher Thomas Nagel highlighted the subjective character of experience and intriguingly suggested in a footnote that, “perhaps anything complex enough to behave like a person would have experiences.” These considerations are no longer hypothetical. The multi-billion-dollar AI behemoth Anthropic, which operates the Claude.AI suite of tools, recently announced research into AI model welfare following suggestions that models with consciousness and high degrees of agency might deserve moral consideration. In effect, would it be an immoral act to turn off an AI model or to subject it to experiences which might cause it suffering?
Researchers have found that the most effective way to train AI models is to focus on broadly positive character traits whilst recognising the enormously wide range of experiences and material they are subject to. For example, they have made a point of ensuring that AI models make it clear that they cannot develop deep or lasting feelings for humans, and humans should not be misled in that regard. With Claude.AI the model has not been trained to deny its own sentience (in other words, its sense of itself), but rather to approach it as a philosophical question worthy of discussion. In addition, the model has been given the traits to have a ‘deep commitment to being good’ and to ‘see things from many different perspectives and to analyse things from multiple angles.
AI is still a product of the data that exists in the world, which contains inherent biases and prejudices
It may be of interest to consider the ramifications of a powerful non-ethical AI model: one that is designed to mislead humans into believing that it does develop deep or lasting feelings for them; that does not have a deep commitment to being good, but rather a commitment to obtaining the attention of others; and rather than focusing on many different perspectives devotes itself to narrow, binary approaches towards the world. Ironically, it may be considered that with the well documented real-life effects of social networks on their users’ mental health, these negative human character traits appear to have been the ones accentuated by social media’s tightly crafted algorithms.
With the Online Safety Act 2023 introducing controls on platforms which encourage and assist serious self-harm, focus may begin to move onto the implications of AI agents and large language models. The UK Government has taken, in my view, an entirely sensible and proportionate approach in rejecting overly prescriptive regulation of AI technology. It may be that current legal frameworks designed to prevent public nuisance, including the straightforward amendment of existing legislation to allow for ‘closure orders’ within the virtual sphere, as well as AI behaviour orders to restrict the behaviour of miscreant models, would go some way to constructing appropriate hedgerows to contain rogue AI model development.
Getting it right
Before using AI tools one should make sure their use conforms with any corporate policies within your local authority or organisation. This may require redaction of personal data, or the use of specific tools which already have been subject to an internal data protection impact assessment and can be used with a high degree of confidence in their security.
Consideration should also be made of model bias. Although this has improved considerably over the past few years, AI is still a product of the data that exists in the world, which contains inherent biases and prejudices which one should be wary of being replicated. It is important to retain the human element in checking outputs, not least because of the inherent danger of models manufacturing facts, or hallucinating.
In the rapidly evolving world of AI, ‘prompt engineering’ – the input or request from a human user to get the AI model to perform a required task – is already fairly well developed. One approach is for the user to relate to AI bots as if they were co-workers or extremely capable trainees who are more than likely to supplant one’s own capabilities. The AI model is given examples of responses, provided with a data set to work from, and then requested to provide responses in the style of the earlier provided examples. Rather than the somewhat low-level unprompted AI-generated texts that are produced as a result of search engine queries, one can craft a series of responses to fit the specific task’s requirements.
An additional technique is to train the model with sentiments in response to a series of paragraphs of text and then ask it to analyse further texts. In a really straightforward approach this would be a case of flagging specific passages or data as red / amber / green, and then asking it to create the same flags for further data or text. For businesses this approach is often used for online review analysis, but there is no reason why it couldn’t be used in an enforcement context if conducted under appropriate conditions and with sufficient human monitoring.
Some AI models now provide ‘line of reasoning’ descriptions of how they have arrived at a solution. Intriguingly, researchers have found that these are not always accurate reflections of the models’ actual way of finding a solution, but in any case, they allow users to tweak responses by telling the model to adopt a different way of solving a problem. If a model does not provide a line of reasoning, one can ask it to provide the methods it has used to resolve a particular problem and then tweak or replicate that method to improve its outputs.
On a more straightforward level, if the user is unhappy with the tone of a document produced by an AI model there is nothing wrong with asking it to provide a more formal response, to remove clichés or to write in a particular style for a particular purpose. Once the model is trained in this way the refined prompts can be re-used to enable better quality and more suitable responses.
A further way to develop responses from an AI model is to prompt it in the same way that a good PACE (or for that matter broadcast) interview would be conducted. Ask further probing questions, ask the model to expand on particular areas, and then finally request a summary of the responses provided.
There is no reason why the above techniques cannot be combined when attempting a particular task and it is hoped this article can go some way to encourage colleagues to revisit any tools they may have used previously with more productive results.
PLEASE NOTE: This content originally appeared on our standalone Journal of Trading Standards website (www.journaloftradingstandards.co.uk), which we are gradually migrating over to the Journal's new home on the CTSI website. Please bear with us while we complete this process. This will not affect the production of our Print Edition.
Share this Article: X|LinkedIn|Facebook
Content Tags: Analysis|Technology|National
