How AI & ChatGPT will shape Tech teams in the future: Embrace it or get left behind…
7th September 2023
In our rapidly evolving technological landscape, artificial intelligence has become a driving force in shaping the way we live, work, and communicate. One prime example of AI’s profound impact is the emergence of ChatGPT and other AI solutions in various industries.
As these advanced technologies become increasingly prevalent, it’s essential to adapt and embrace them. Otherwise, we risk being left behind.
With every new technology though, business leaders must proceed with eyes wide open, because the technology today presents many ethical and practical challenges.
Our expert speaker panel shared their thoughts on the benefits and risk of adopting AI within teams going forwards, and took part in a question and answer session.
Chris Cahill from AND Digital hosted the session and was joined on the panel by Louis Brook (CIO at Lowell Group), Stuart Mills (Assistant Professor of Economics at the University of Leeds) and Dr Antesar Shabut (Senior lecturer in Computer Science Lead at Leeds Trinity University).
Q+A
What is your definition of AI?
Louis Brook: AI is nothing new – and we’ve been sold ‘AI’ for the past 10 years. However, the proliferation large language models we’re seeing real excitement about.
Dr Antesar Shabut: Any computer system that can mimic human behaviour. Any intelligence system that can do useful things for us and improve knowledge.
James O’Toole: Most people’s first experience of AI is of a chatbot but these are programmed models, not true AI. More recently large language models (LLM) are seen more – these are quite different. But there is no code to inspect, just weighting. This AI rather than programmed models is transformational.
Stuart Mills: There are 25 different European definitions of AI. The Deep Blue chess computer was cutting edge AI in its day, now a chess app on your phone can beat you. AI produces output that is intuitive to a human. This definition is helpful in its broadness.
ChatGPT has hit the popular consciousness, everyone has heard of it. Why is this?
SM: The biggest advance here has not been the AI but the UI, the user interface. Coders could interact with AI before but now non-coders can, this is a behavioural change rather than an AI change.
JO: OpenAI put out instructions a few years ago about how to interface with AI – but nobody knew about it until the nice user=friendly UI was added. AI is a mirror for humans, we imagine ourselves dealing with a person. We like ourselves and if machines are like us then they will be popular.
AS: A few years ago nobody knew how to use AI save for a few techies. Now anybody can use it. There are lots of open resources available so it is simple to develop an AI model. The most difficult part is to optimize and implement for real life. The availability of a simple UI is new, for accessing AI.
LB: Accessibility of AI especially to a model primed with datasets is mind-blowing for people. The question now is how we make use of this.
What are the benefits of AI?…..Good and bad use cases?
LB: A use case I’m looking at is around decisioning. An extension of machine learning, spotting trigger events for outcomes.
Easy automation of certain channels – intent / sentiment understanding.
The benefit is usually around efficiencies – headcount etc. and enablement – development is obvious one also copy writing and legal scripts.
AS: For me, mainly research so far, it is good for looking for materials and for lesson planning. Also editing and improving plus proofing writing. It can be used for brainstorming and helps focus on ideas and finding solutions. It can be helpful as a recruitment tool and for interviews.
JO: There are far too many use cases to list but the legal point is apt, as small business it can be useful for employment contracts for example for tax, for GDPR, it can save you professional fees.
It is useful for development – especially for junior developers who can use it to pair with, asking it questions and helping with debugging and syntax. It can certainly help to upskill junior developers.
The biggest use is for scaling conversations with customers – for example sorting and prioritizing huge numbers of inbound emails. Also for customer service – how to fix things. And prioritizing and categorizing, if not able to answer the question.
SM: On the subject of unlocking access to otherwise expensive professional services, the question will be one of the quality of the information provided. Going forward the main use will be for information processing and decision making. It can handle a deluge of potentially useful information that humans would struggle to process. Of course it is possible to bespoke AI models to enhance the quality of the decision making.
There has been much in the press about plagiarism in academia, then even using the same tools to detect it – what is the current position?
SM: Everyone’s head set on fire when ChatGPT was launched. The flames are beginning to subside now and some clarity emerging. Many universities banned ChatGPT altogether although in many high schools (anecdotally, I have no data around this) is it being massively used although there may be issues around quality. My own students are much better when they write their own answers and when they use it, I can always spot it and I wonder, what does it say about their own self confidence and self-belief? The driver for plagiarism is the real concern, not the plagiarism itself.
JO: People worried in a similar way when the internet started and especially when Wikipedia was new. These things can be a teacher and that can be more valuable and outweigh the risk and danger they present.
AS: Most institutions embrace this tool for the benefit of their students. All universities have policies around this and training in how to use it ethically. It can be okay to use it as a tool but for it to do the work for you. It is important to utilize it in the right way.
LB: Thinking about industry, most developers already code by plagiarism – from Stack Overflow and elsewhere! What’s the difference? This is just the next generation of how we make computers do stuff. We should foster innovation and GenAI is no different from anything else in that there needs to be governance around things going live to ensure resilience and stability
This begs other questions; if it “wrote itself”, how do we know how our decisions are made? This may have bad outcomes for customers. Boards are now setting up IT Ethics Committees to look at this.
So once you have your use case, how should you start your AI journey?
LB: Wait for others to go first and see what happens!
But think about the future; applications without documentation are already difficult to support and potentially becomes very dangerous if you’re relying on AI to build your product.
AS: Consider your team, decide what you want to achieve from AI, match it to your budget and your team.
JO: OpenAI is the most popular route to take but there are many other tools out there. Go with OpenAI because it has been prominent in the field for the longest and has not just been jumping on the bandwagon. There is concern around bias and safety –OpenAI shines in this area and has taken steps to address the issue. It is also good at not being rude! Secondary to safety andbias is GDPR and privacy. If deletion is requested, then you must delete so if you train you model on customer data, you mayneed to “un-train” it to make it forget. Best not to roll your own – choose a platform and I would recommend OpenAI becausethis is all addressed by the OpenAI APIs.
SM: Long term, this is like choosing your accounting package. Shorter term, it is a management decision about allowing employees to play. So be sensible, don’t give the system the company credit card details!
In terms of societal impact, what can we expect to see in the next 12-24 months?
SM: 2017 – a famous paper was published predicting job losses from AI.
2019 – the ONS replicated this saying that the knowledge economy would be fine, the lower skill would be fine but it was those occupying the middle ground who would be vulnerable. Roles using the brain a bit but not as a core part of the role was where the danger lay. (They included accountants in this!)
Now, 5 years on, more knowledge work seems automatable. Usually this sort of thing creates more jobs than it eliminates. The big issue will be the quality of the jobs and of the automation too.
JO: Beware over-hype and losing headcount who solve problems – if you lose your critical knowledge base it is hard to build up again.
AI will not take over the world but certain sectors will inevitably be affected – translating, copy writing, developing perhaps?
AS: I am a tutor and I was worried! Seriously though, AI will bring different jobs to society – we all know about the digital skills gap, students must meet employers’ needs. We must keep up to date with digital transformation and keep developing the skills needed for these jobs.
LB: There is nothing new about automating customer interaction – self-service this is a direction of travel in this context and AI will play a role here. More widely, I see positive impacts for individuals and SMEs where there is now access to capability and knowledge that would previously be costly.
How can how to use AI to help with collating and curating a tender library?
JO: Definitely use ChatGPT to help filter down bid responses if you are reviewing tenders to prioritise and score against your metrics. Learn how to make prompts to help it to help you. Use it to summarise for example, you do not need to be a super-expert to do this sort of thing.
SM: Anecdotally, an academic consulting business used to be content with winning 4 bids per year have won 20 since they started using ChatGPT – possibly because it is massively easier to put out more bids – and query, are they better quality?
Tendering will become more competitive as more bids are received – so bids need to be better!
LB: Spend more time understanding what makes a good bid and outstanding bid – differentiate in this way.
AS: Use the basic paid version of ChapGPT – the open version’s information stops at 2021.
Would you advise an organisation’s tech team to embrace GenAI wholesale or, given that we are on the high of a hype curve, would you advise watching and waiting?
LB: Governance comes first, non-functional requirement don’t go out of the window Do not ignore this. But you can embrace and limit exposure – go at a safe pace.
SM: There should be concern about going too fast because it can lead to mistakes. The computer revolution in the 1980s was massive but it took decades for real profitability from it to filter through. Where the real value of AI lies will take time to filter through too – fear of being left behind can lead people to go too fast.
ChatGPT is not human, can it really mimic a person?
JO: Fine tuning is the answer here, take a large language model and add a few layers of training, a corpus of example data, it will learn to sound like you. There are limitations for sure but it can be trained to sound a certain way.
AS: A survey in a US university found that students could not tell machine from human!
LB: If you simply ask ChatGPT to adopt a particular style it can do this – try it!
How do you think GenAI will cope with rules changes such as to GDPR or the law / regulations?
LB: To avoid the consequences of GDPR non-compliance now or in the future, use data for sure but annonymise it before it is used or learned from.
SM: I am speaking at a conference shortly about regulating behavioural technology. AI is behavioural technology. Regulatory bodies are very concerned about AI, especially in the financial world. Hard rules-based regulation is no longer adequate. Principles-based regulation is better to ensure non exploitative and equitable deployment of AI.
How do we avoid AI becoming biased over time, i.e. it has been trained properly at the outset but data use over time changes this?
AS: You cannot de-risk that but can mitigate it – be careful of the data quality and the amount of it – it is not just the developers who are responsible for bias in a model, it is the whole team.
LB: Don’t train on biased data of course, but this is one for those new IT Ethics Committees!
JO: There are tests available on data in vs data out.
And of course, machines are a mirror in that if you do not like what you see, don’t break the mirror, change what it is reflecting! The machine will reflect what it sees.
SM: Where algorithms are used in UK Government policy making, humans are required to in the loop, for this very reason.
It is a very interesting challenge where quality is degraded by AI training AI…
How concerned should we be about our clients’ data privacy – how robust / reliable / secure are AI tools? ChatGPT is not an enterprise-level solution – should we be worried?
JO: You should be very concerned. GDPR exists and influences decision. Authentication of users of ChatGPT has been notably lacklustre – access to the API key is being addressed now. It is an education issue as well, not putting personal data into the ChatGPT web version.
LB: These non-functional requirements will never go away and are absolutely paramount. This is no different from the governance around any SaaS or other tool you might use.
AI hallucinations – can you train these out?
JO: AI hallucinations (I am not a fan of the word but it has stuck) are sometimes called lying but that connotes knowledge of doing so – not the case here. But non-truths, or making things up, is definitely an issue.
The solution is to give more context, extra information. Where this is not possible, prompt to say “don’t know” or to hand-off to a human. Fine-tuning will help, start with a prompted model then correct / re-train to remove hallucinations. But of course you need the answers for this. One of the dangers of over-hype around AI is that plausible answers may be taken as correct but be completely wrong.
SM: I don’t like the term either, it implies AI is thinking and undermines the design of the engineer.
Hallucinations are blank-fillers for gaps which engineers have presumed consumers will want to have filled. It is not an error, it is a conscious design decision.
Trust – my team do not trust the large language models, they feel dangerous because you cannot extract back, your data could be leaked or exploited by other agents, what do you think?
LB: Read the privacy policy of your supplier very carefully to see the use to which your data is put.
AS: Models are often called “black box” which simply means we do not know what is going on inside. You cannot trust everything so do not enter sensitive data into large language models.
JO: Remember the Samsung example when an engineer using the web interface for ChatGPT entered data about CPU processors and someone found it! OpenAI have now mitigated this risk but it is still a significant risk, and OpenAI are transparent about training their model with data entered on the web interface. For APIs, data is encrypted and kept for 30 days for inspection purposes then it is deleted. It is never used for training the model – so they say anyway! The greater risk is around training in-house models.
Speakers
Louis Brook
Louis Brook is the CIO for Lowell Group, one of Europe’s largest credit management services providers. Louis has 20 years’ experience within financial services technology having spent much of his career with HSBC Group, most recently as CIO for the Leeds-based digital subsidiary bank, first direct. Louis’ passion lies in exploiting tech for good, and thrives in inspiring teams to deliver solutions that empower consumers to make better financial decisions.
Stuart Mills
Stuart Mills is currently an Assistant Professor of Economics at the University of Leeds. His research focuses on economic policy surrounding behavioural science and artificial intelligence. He has spoken on these topics for organisations such as the United Nations and the Behavioural Insights Team. He is the author of Artificial Intelligence for Behavioural Science, and has published several articles on the intersection of AI and behavioural science. His work has influenced policymaking at the local and national level, and he is currently leading a project on the applications of Behavioural Science for local economic development.
Dr Antesar Shabut
Dr Antesar Shabut is a senior lecturer in Computer Science and CompSci Programme Lead at Leeds Trinity University. Dr Shabut’s current research interests lie in the areas of applied AI to mobile-based intelligent systems, image processing, computer vision, and trust and cyber security modelling in distributed systems. She has published several papers in reputed journals and conferences. She is also a member of Women in AI and Teens in AI groups, and she has been involved in different activities that support digital diversity, equality and inclusion in HE and the workplace.
Chris Cahill
Chris has been with AND Digital for nearly five years in a variety of Tech leadership roles. Prior to that Chris, was at Hermes and Yorkshire Water.
Chris does a lot of work in the Tech community in Leeds, having been involved with the British Computer Society for many years in a range of roles including Treasurer and Chair, and also as Regional Chair (North and Midlands) of BCS ELITE.
Future Events
For information on upcoming 4it tech events, please email us at sam.devonshire@4itrecruitment.co.uk