AI in the workplace: a whistle-stop tour for employers

Transcript

Please note this transcript is an automatically generated summary and may contain inaccuracies.

**Katherine Cooke:** Good morning everyone. Thanks for joining me this morning. My name's Katherine Cooke, I'm a senior associate at Higgs LLP, and I'm delivering a talk this morning about AI in your workplace. Now, this might well be something that you've seen hit the headlines for a variety of different reasons, and it might be something that your employees are starting to use as well. This is just a little bit of a whistle-stop tour of AI, how it might affect your workplace, and things you might start to put in place now to establish some rules and regulations about how you use AI in the workplace and what's appropriate use or isn't.

What is AI?

**Katherine Cooke:** So first of all, what is AI? It's technology that enables computers and machines to simulate human intelligence and problem-solving capabilities. That's not my definition, that's what I've borrowed from IBM, but it's my emphasis there on the word "simulate". I think what sometimes gets lost in all the noise and publicity about AI is that this is not something that can generally think for itself.

Although you'd be forgiven for thinking that if you are having an interaction with functions like ChatGPT, for example, it's really difficult not to feel that you're having a two-way conversation with something. But you have to remind yourself that it's not a person, it's a simulation of a conversation. It is a program that is reactive to certain prompts and mining data to try and give you the answer that you wanted and to present it in a certain way. There is no critical thinking involved from certain AI programs, no fact-checking, and no appreciation or evaluation of the sources of information that the program might well be using.

It can be used on its own or combined with other technologies such as sensors, geolocation, and robotics, and it can perform tasks which would otherwise require human intelligence or intervention. It's also sometimes mentioned alongside machine learning and deep learning. And again, that kind of deep learning also includes the development of AI algorithms, modelling decision processes of a human brain, and it can learn (I use that with inverted commas again) from available data and make increasingly more accurate classifications and predictions over time.

Practical examples of AI

**Katherine Cooke:** So some practical examples of where you might see AI already in your day-to-day. Quite often people have chatbots on their websites that come as a kind of little customer service speech bubble that pops up saying, "How can I help?" It is usually an AI chatbot, speech recognition. That's something that we at Higgs use as well in terms of dictating notes of meetings and things like that. And again, that might all be something you've come across in your day-to-day activities or again as an adjustment for an employee who might find typing more difficult because of a disability.

Computer vision is another application. It allows computers and systems to take information from digital images and to take action as a result. And again, that's lent itself really to applications in processing imaging from healthcare. So looking at scans to determine if this is a scan of a healthy person or if there is a cause for concern here that needs further investigation.

It's also used in supply chain analytics, so they can use algorithms to analyse data about how many products they ship, where they end up, et cetera. And weather forecasting, rightly or wrongly, usually uses algorithm modelling as well.

The rise of large language models (LLMs)

**Katherine Cooke:** Really in 2023, AI became hugely topical because of the rise in a particular type of AI called large language models or LLMs. These are programs like Chat GPT, and with these new generative AI practices and deep learning models that can pre-train on these big amounts of data, then there needs to be some assessment of how these might work for your workplace and might need to be utilised to make you more efficient, but also where there might be some potential traps, teething problems, and issues as there always is when we're implementing something new.

What are LLMs?

**Katherine Cooke:** So LLMs, as I've said, are computer programs that can recognise and interpret human language or other types of complex data. Many LLMs are trained on data that's been gathered from the internet. Some models learn from more closed data sets, which are more focused, and they also learn from the data that you insert into them.

So, in certain cases, the free AI tools, you are the product. In some cases, you are helping that algorithm learn. Just a few little names that you might come across of the different types of LLMs. As I say, some are more well-known than others. It's probably not the scope of this talk to go through all of them and their differences, but again, just a kind of general awareness piece that there are lots of different models out there that are freely accessible and that you may want to be aware of.

Potential uses of LLMs in the workplace

**Katherine Cooke:** So what are the potential uses of LLMs? How might it be helpful in your workplace? Certainly, we've seen it's helpful in drafting articles. If you have content for website publicity purposes, articles, all that kind of thing, AI might help you with that tricky first sentence of getting off the ground of what you want to say, how you want to say it, and what content you want to put in.

Likewise, it might help you in developing codes for software and other functions. It can help generate images. I know that's been a huge topic of conversation as to the veracity of AI-generated images and how effective they are. Sometimes there are common problems, particularly around hands, and accurate generation of hands and sometimes faces and facial expressions, but again, that might all be something and problems and issues which are ironed out over time. It can be used to solve maths problems. It can also answer general questions and, as I say, it can generate content.

I think we can perhaps begin to see a little bit of where AI might well be useful in your business and actually why it might be attractive to some employees to utilise those three AI tools to help them in their day-to-day work.

Potential problems & issues

**Katherine Cooke:** But, and there's always a but here, what are the potential problems that you might need to be aware of, particularly from an employment law perspective?

CVs and the recruitment process

**Katherine Cooke:** The first one I can think of really is CVs in the recruitment process. How often is it that you have an applicant who looks amazing on paper, but if you get them to an interview or get them to carry out an exercise independently and supervised, or in fact at the point at which they start doing the job and their boots are on the ground with you, that they are not really the applicant that they sold themselves as being, that there's a bit of a disparity between the abilities that were shown in the recruitment process and what you can see in your day-to-day assessment of their performance.

In the olden days, I suspect that was probably because somebody had helped them with their application or reviewed it or amended it. These days it might well be that somebody has had a little help from Chat GPT or other LLM models in terms of their written comprehension.

So how do you deal with that risk in your recruitment process? You might think about updating your recruitment policies and materials to set up very clearly what the consequences will be for potential applicants if it is found that they're using AI in your processes or to answer questions or carry out surveys and other bits and pieces that you might have as part of your assessment processes. It might also be good to consider changing your process so there is a bit of a degree of supervision. There is a degree of on-site work assessment which can be carried out and you can be confident that it's the applicant's own work. Obviously, this is something that schools will grapple with for a long time in terms of exam coursework and appropriate supervision. And again, I'm sure they're facing incredibly similar problems with the use of Chapter P to carry out homework and carry out coursework recruitment decisions.

Bias in AI-assisted CV screening

**Katherine Cooke:** One of the ways again that AI can assist businesses is screening CVs and trying to match good applicants with your vacancies. The AI can scan those applications and rank them according to predetermined criteria. For example, screening out anybody who's got less than a two-one or looking for a particular skillset or particular qualification, or it can be used to look for keywords or key concepts.

However, again, this can go wrong. Amazon is the example I'm going to be using here. They implemented an automated recruitment tool that had utterly resulted in a discriminatory bias towards female applicants. They created the algorithm, it was perfectly neutral, it was to find good applicants for jobs regardless of gender, age, et cetera. But the CV screening system self-modified itself and preferred male candidates. This was in part because the data it was training itself on, Amazon's historical applicant data, pretty much the majority of those successful applicants had been male. So it concluded that it should downgrade any sort of women's application in the system, and it was downgraded by the algorithm.

So perhaps a classic example of, "I'm not paranoid, the machines genuinely are working against me." The AI is only sometimes as good as its input from the data it's accessing, and it might well regurgitate bad decisions and biases which might be inherent in that data that you've gathered. So again, a word of warning there that it might still be an issue for a human eye to look over shortlisted applicants if you are applying AI in your business in that way.

Also, again, particularly if you have long-term aims like closing the gender pay gap, for example, or widening your recruitment pool to encourage applications from people of wider, different backgrounds, then that's potentially going to be problematic, and you would need to set up your AI and algorithm to assist you with that. As I say, have somebody who can actually assess what the results are that it's producing.

Discrimination in work allocation

**Katherine Cooke:** Likewise, AI can also be used in work allocation. The practical example of this is Uber. The drivers are required to log into an app, say they're ready for work and ask what jobs are available. App access was through facial recognition.

The trouble with this facial recognition on the app was that the authentication software had greater difficulty recognising dark-skinned faces, which resulted in those users being unable to access the app and find jobs. So you can immediately see there's disparity if a white colleague can log in, take the nicest available jobs or the most lucrative available jobs, but you are stuck not willing to access the app because it can't recognise your skin as easily.

The disparity was more apparent for dark-skinned women, with a failure rate of 20.8% compared to a failure rate for logging of 6% for males. So again, a slightly troubling issue with software development there, and again, that it might lead to consequences which are not necessarily anticipated.

Intellectual property and content creation

**Katherine Cooke:** As I've said, AI can be a really useful tool for generating publicity and other content, and lots of people who might not feel comfortable with written communication for whatever reason, or that their strengths are not particularly strong in producing it, might use Chat GPT as a useful way to save time and improve their output.

So I suppose the question is, who does that material belong to? The model for Chat GPT is trained on creative works, and there are at least in America to my knowledge, there's at least one case proceeding where some authors, a collective suing Open AI (who are the makers of Chat GPT) over related use of their work because they say, actually, if you are training this model on the work that we produced, or you are taking something which is our intellectual property and you are gaining a benefit from it that we are not seeing, and that gives unlawful, we wait to see what the outcome is of that case, but there are likely to be legal actions over here and elsewhere on that basis.

Other content providers, I've had client experiences with Getty, who found images on our website which belonged to Getty, and Getty were quite vociferous about taking that image down and not using it and being very protective of their work. Again, I can foresee that being a problem. If you are producing your own content with the help of chat, you can see there might all be an issue of who owns it, particularly if you are using that content is something you sell on or license to your consumers or clients.

So yeah, be very aware if an artist can recognise some of their content in your content, that could lead to difficulties, problems, potential legal action, et cetera.

HR management

**Katherine Cooke:** HR management is also another area where you might consider farming out some elements to AI. For example, answering employee queries, using data collection to improve performance assessments. Particularly if you have ways and means of monitoring your employees, for example, piecework time recording, other output measures. It can also identify employees who require additional support, and perhaps again, in terms of meetings that you might have for grievances, disciplinaries all sorts of the processes, AI can assist in helping you automatically transcribe and record those employee meetings. It can also make decisions about grievances, hiring disciplinary procedures based on employee data.