Risks and considerations

Introduction

If AI is the wonderful thing that people claim it to be, what’s the problem? Well, there are actually a whole lot of problems and risks which you have to consider as part of your due diligence and this is where ethics comes in.

In this context, ethics is the systematic practice of distinguishing right from wrong in decision-making, moving beyond legal compliance to address the broader moral implications of technology. It means actively questioning the impact of AI on human dignity, societal fairness, and ecological stability, rather than accepting technological capability as a justification for deployment. For organisations, this requires a shift in perspective: viewing ethics not as a barrier to innovation, but as the essential compass that guides responsible development and ensures that progress does not come at the expense of fundamental values.

So, in order for an organisation to “do it right” or “do it responsibly”, ethics must be the primary lens through which all other risks are viewed. Whether assessing the environmental footprint of data centres, the labour conditions within the supply chain, or the potential for algorithmic bias, an ethical perspective ensures that these issues are not treated as isolated compliance hurdles but as interconnected elements of a broader responsibility. By embedding this mindset into strategic planning, leaders can proactively identify and mitigate harms before they occur.

The following sections outline the critical areas where this scrutiny is most vital, providing a foundation for organisations to navigate AI adoption with integrity and foresight.

 

Risk of Bias and Discrimination

AI systems are fundamentally data-driven, and the outputs produced through AI systems are always a reflection of the data that those systems were trained on. Yet data is often a reductive and incomplete representation of reality. It captures what has been measured, frequently omitting the nuance and complexity of human experience and encoding historical prejudices. When models are trained on such datasets without intervention, they risk perpetuating or amplifying discrimination based on race, gender, age, or socioeconomic status. This "bias in, bias out" dynamic transforms past injustices into automated future decisions. For example, AI systems used in recruitment have often led to male applicants being identified as the most suitable candidates, because historic data used to train the model reflected previous prejudices of human recruiters. Presenting AI decisions as being mathematically objective or value free can make discriminatory outcomes harder to detect and challenge than human bias.

The risk extends beyond who is represented to who is missing. Gaps in data collection can render certain groups invisible, while flawed metrics may optimise for proxies that correlate with protected characteristics. In high-stakes areas like recruitment or talent identification, this can systematically exclude marginalised groups and undermine fairness. Mitigating these risks demands more than technical fixes; it requires a critical contextualisation of data. Organisations must actively audit datasets for representation gaps, question the historical context of their information, and maintain rigorous human oversight to ensure AI tools promote inclusion rather than cementing past prejudices.

 

Environmental Impact

While it often seems like digital technologies like AI are somewhat abstract and do not have a physical presence in the world, in reality their environmental impacts are substantial.

The environmental footprint of artificial intelligence extends far beyond the electricity used to power data centres; it is embedded deeply into the supply chain, development, and operational use of the technology. At the extraction stage, the mining of critical minerals such as lithium, cobalt, copper, and rare earth elements causes profound physical landscape destruction. Open-pit mining and chemical leaching processes strip vast areas of vegetation, destabilise soil structures, and generate toxic tailings that can contaminate local water tables and ecosystems for decades. This upstream impact is particularly severe because the rapid obsolescence of AI hardware accelerates the demand for new raw materials, creating a cycle of continuous extraction that outpaces natural regeneration and sustainable land management. Electronic waste, or ‘e-waste’, (created by the obsolete hardware) is also one of the fastest growing waste streams.

During the development and manufacturing phases, the environmental cost shifts to energy-intensive processing and fabrication. Refining rare earth elements and producing advanced semiconductors require immense amounts of electricity and water, often in regions where the power grid relies heavily on fossil fuels. The chemical processes involved in chip manufacturing also release hazardous pollutants and greenhouse gases, contributing to local air quality degradation and global climate change. Furthermore, the global logistics network required to transport these components from mines to refineries, then to factories, and finally to data centres adds a significant layer of carbon emissions through shipping and air freight, compounding the total lifecycle footprint before the AI system is even switched on.

In the operational phase, the environmental burden manifests through sustained high energy consumption and water usage, sparking growing controversy and local resistance. Training large-scale models and running inference tasks require massive data centres that generate intense heat, necessitating cooling systems that often consume millions of litres of fresh water daily. This has led to significant pushback in various regions, including the UK, parts of Europe, the United States, and in South America (see the AI Resist List) where communities and local authorities are increasingly opposing the construction of new facilities. Critics argue that these centres strain local power grids, divert scarce water resources from agriculture and residential use during droughts, and disrupt local ecosystems. As a result, permitting processes are becoming more contentious, with some projects facing delays or cancellations due to public outcry over their unsustainable resource demands and lack of transparency regarding their long-term environmental impact. So, when the technology sector talks about digital and the “cloud” we must remember that what lies beneath is not ethereal but is very real and physical.

When it comes to individuals’ own use of generative AI tools, environmental considerations should be one of the factors informing decisions about which tools to use, and for what purposes. Accurate figures about the actual impacts are difficult to obtain due to a lack of transparency from AI companies, however one study estimated that for every typical conversation with ChatGPT the model used 500ml of water. Another study estimated that created a single image using a generative AI model has the same carbon emissions as fully charging a mobile phone.

More on sport and the environment

 

Human Rights and Labour Exploitation

The human rights implications of the AI supply chain are significant, beginning with the extraction of critical minerals in regions where regulatory oversight is often insufficient. In areas such as the Democratic Republic of the Congo, mining operations for cobalt and other essential materials have been consistently linked to serious labour violations, including the use of child labour and hazardous working conditions. These practices contribute to local instability and community displacement, creating a stark disparity where the economic benefits of AI technology are largely realised in developed nations, whilst the social costs are borne by vulnerable populations in resource-rich regions. Despite international attention, mechanisms for protecting affected communities remain weak, and the opacity of global supply chains continues to hinder effective enforcement of ethical sourcing standards.

Labour exploitation also extends to the digital workforce essential for training AI models. The development of sophisticated algorithms relies heavily on data labelling and content moderation, tasks frequently outsourced to workers in lower-income countries. These individuals often operate under precarious employment conditions, characterised by short-term contracts, low piece-rate pay, and a lack of basic benefits or job security. In addition to precarious working conditions, content moderators tasked with filtering harmful material for AI training data are routinely exposed to disturbing text and imagery, leading to documented psychological distress without adequate professional support. While this workforce is fundamental to the functionality of modern AI systems, it remains largely invisible in public discourse, underscoring an urgent need for greater transparency and robust labour protections across the industry.

More on sport and supply chains

Explainability, transparency and accountability

The opacity of advanced AI systems, often called "black boxes," challenges organisational governance. True transparency extends beyond technical explainability, i.e. just the understanding of the algorithmic mechanics, to include the human decisions surrounding the system’s intent, scope, and limitations. Without this holistic clarity, organisations risk eroding trust and failing in their duty of care.

Many high-performing models are inherently difficult to interpret. Relying on such systems for high-stakes decisions without a clear rationale creates an "accountability gap,"(when it is unclear who is responsible for an AI-driven decision or its consequences, leaving no specific individual or entity liable for errors, bias, or harm) undermining fairness and exposing organisations to legal challenges under regulations like the UK GDPR. Furthermore, transparency should cover the context of use - organisations must explicitly state why AI is deployed, disclose its specific limitations and data sources, and avoid hiding biases or constraints behind technical jargon. Concealing the "why" or the data provenance (i.e. where it comes from and its authenticity) prevents stakeholders from assessing the ethical validity of the project.

But transparency is meaningless without clear accountability. A common failure is the "rubber-stamp" human-in-the-loop, where staff nominally review AI outputs but lack the authority or time to override them. This creates a facade of oversight where humans are blamed for algorithmic errors they could not prevent. Organisations must define a clear chain of responsibility: a named individual or body must be empowered to audit the system, interpret its outputs, and accept liability for its consequences. Accountability cannot be outsourced to vendors or algorithms. Additionally, explanations must be accessible as publishing dense technical reports that non-experts cannot understand is merely "ethics-washing," and not true transparency.

In sport, where fairness and integrity are foundational, explainability is critical. Decisions on athlete eligibility, discipline, or officiating must be defensible and open to challenge. An AI system influencing these outcomes must provide a clear, non-technical rationale. Sports bodies must also be transparent about training data, acknowledging any historical biases that could skew talent identification or performance analysis. Ultimately, maintaining the legitimacy of sport requires that technology acts as a transparent tool for justice, with clear human accountability, rather than an unquestioned final authority.

 

Risk of misinformation and disinformation

Misinformation and disinformation both refer to false or inaccurate information, but what is the difference? The difference lies in the intent:

AI can accelerate misinformation and disinformation by making fake content easier to create, harder to detect and quicker to spread.

Generative AI tools allow anyone to produce convincing text, cloned voices, and deepfake videos at almost no cost. During various recent political elections, these technologies were used to spread fabricated narratives rapidly. AI algorithms on social media also amplify this content by prioritising engagement, often boosting sensational lies over nuanced truths.

Deepfake content affects not just high-profile public figures but everyday people. More than 90 per cent of deepfakes produced are pornographic in nature. The creation and sharing of non-consensual sexual imagery are major problems that disproportionately affect women and girls. AI generated images and videos are increasingly being created to bully, humiliate, degrade and even blackmail individuals.

While AI fuels the problem, it can also be used to detect fakes and assist fact-checkers. However, the sheer volume of realistic synthetic media has eroded public trust, allowing bad actors to dismiss genuine evidence as AI-generated.

The spread of false information creates practical challenges for society by making it harder for people to agree on basic facts. When trust in news and institutions falls, communities struggle to work together on common issues like public health or environmental policies. This often leads to greater division, as groups rely on different sources of information and find it difficult to have productive conversations.

For democracy – and this includes, for example, internal electoral processes in member organisations – the main risk is that voters may make decisions based on inaccurate details, while elected officials find it easier to dismiss valid criticism by claiming it is fake. Over time, this confusion can cause people to feel tired of the news cycle and less interested in participating in civic life. While this does not mean truth has disappeared, it does make the process of finding and agreeing on it more difficult for everyone.

 

Risk of hallucinations and inaccurate outputs

Hallucinations is a term that is used to describe responses generated by AI (primarily LLMs) that contain false or inaccurate information presented as fact.

But it is a problematic term as it anthropomorphises the software and suggests that AI models normally know what is real with hallucinations representing instances where something has gone wrong. Ultimately, LLMs do not fundamentally understand what words mean, nor have any concept of what is true or real. As such implying that an AI model is somehow hallucinating is incorrect and leads to people equating a human cognitive process with statistical pattern completion.

It is literally an LLM’s function to pattern match the most statistically probable words. For the functioning of the LLM, the question of whether those words are correct or not is completely irrelevant. This is why you must be wary that inaccurate outputs are highly probable, especially if prompting for answers to uncommon questions or situations as there will have been less data available to train the model.

Therefore, framing hallucinations as a problem to be 'fixed' is misleading, as LLMs are functioning exactly as designed. Rather it is essential to check the outputs of LLMs for accuracy: a potential duplication of effort that may negate any perceived productivity savings.

 

Data privacy risks

The use of AI carries data privacy risks for all organisations, mostly centring on the potential loss of control over sensitive and personal data. When publicly available versions of ChatGPT and Gemini are used, and users are not sufficiently trained or aware of the privacy risks, this can expose organisations to data leakages where inputs such as client data or commercially sensitive IP can be retained and used for retraining or exposed to unauthorised parties.

UK GDPR and EU GDPR compliance can be complicated when it comes to the use of tools whose infrastructure lie outside these jurisdictions or if the data transfer crosses between compliant and non-compliant territories. And some AI tools are very opaque about their data storage.

AI systems can also be vulnerable to inference attacks, where bad actors can repeatedly query a model to reveal private data even though direct data access is blocked. Beyond simple repetition, attackers often employ sophisticated social engineering tactics within their prompts to bypass safety filters. They may frame requests as urgent security audits, role-play as authorised administrators, or use "jailbreaking" techniques that embed malicious instructions within seemingly benign contexts like creative writing or hypothetical scenarios. By exploiting the model's tendency to be helpful or its inability to distinguish between a genuine user and a simulated threat, these actors manipulate the AI into reconstructing sensitive information piece by piece, effectively inferring the private data without ever triggering direct access alarms.

In the sport and physical activity sector which could involve the handling of health and performance data of athletes and participants at all levels, data privacy risks cannot be underestimated. The sector processes sensitive "special category" data, including biometrics and health records. Risks include repurposing performance data for commercial AI training without explicit consent, failing to explain automated decisions affecting careers (the "black box" issue), and the technical inability to fully delete an athlete’s data from a model’s weights upon request (once it’s out there, it’s out there).

To mitigate these risks, organisations should adopt a "data minimisation" approach, collecting only the information strictly necessary for immediate performance goals and excluding broad historical health data unless explicitly required for medical safety. Implementing strict access controls ensures that sensitive biometrics are visible only to essential medical and coaching staff, while anonymising or aggregating data for general analysis prevents the identification of individual athletes. In addition, establishing clear retention policies to automatically delete or archive data once an athlete leaves a programme ensures that sensitive records do not accumulate unnecessarily, effectively excluding them from long-term storage where they could become vulnerable to breaches. This is good data management practice.

But the need to ensure data security is not simply for organisations with elite athletes on talent pathways. Organisations of all types collect personal data about their members, participants, customers and others. This might include banking details, addresses and contact information, or details of contractual arrangements with suppliers and customers.

It is important to note that if free versions of generative AI models are used, any information put into the system may be retained and the user, or any person to whom that information relates, loses control over how it might be used in the future. This can include commercially sensitive information (budgets and financial data, business plans, product information, etc) or an organisation’s intellectual property. Therefore, considerable caution is needed when engaging with these tools. And indeed, the terms and conditions of paid-for platforms should be checked carefully as to what use is made of data entered onto them.

Cybersecurity and Technical risks

The integration of AI introduces distinct technical vulnerabilities that differ from traditional software risks. AI systems are probabilistic and data-dependent, creating new attack routes that will require specific security measures to ensure integrity and availability.

AI models interacting with users are susceptible to "prompt injection," where malicious inputs trick the system into bypassing safety protocols or revealing sensitive data. Unlike standard cyberattacks, these are semantic manipulations that traditional firewalls may not detect.

AI performance relies entirely on the integrity of its training data. "Data poisoning" occurs when attackers introduce corrupted or malicious data into the learning pipeline, potentially degrading model accuracy or creating hidden "backdoors." While inference attacks (discussed previously) focus on extracting data from a finished model, data poisoning targets the foundation itself, compromising the system before it is even deployed. For sports organisations, this could manifest subtly: an unscrupulous opponent might inject falsified injury recovery timelines or manipulated biometric baselines into public datasets that a model scrapes for training, leading the AI to recommend unsafe training loads or misdiagnose fatigue.

An often-cited example of data poisoning is Microsoft’s 2016 "Tay" chatbot. Co-ordinated users fed the bot inflammatory content on Twitter, causing it to adopt toxic behaviour within 24 hours. This demonstrated how quickly unverified external data can weaponise an AI’s output.

Validated research confirms this threat persists in advanced models. A Nature Medicine study published in 2025 showed that inserting a tiny fraction of corrupted records caused medical AIs to make dangerous diagnostic errors.

Sophisticated actors can also use "model inversion" techniques to reverse-engineer an AI system, querying it repeatedly to reconstruct sensitive information about the individuals in its training data. This challenges the assumption that aggregated data is anonymous. Strict data minimisation principles and privacy-preserving techniques, such as differential privacy, are necessary to protect sensitive information from being extracted via the model itself.

Most AI deployments rely on third-party APIs (Application Programming Interface, which allow software platforms to communicate and share data), open-source libraries, and cloud providers, creating a complex supply chain. A vulnerability in an underlying component or a breach at a vendor level can compromise the entire system. Furthermore, reliance on external services introduces operational risk if a provider discontinues a product or changes terms unexpectedly. Organisations should think carefully before integrating an AI model into core functions or services in ways that would make it difficult to operate if the terms of access to the model were to change, or if there was a security breach.

AI models are not static. Their accuracy can degrade over time as real-world conditions change, a phenomenon known as "model drift." A system trained on historical data may become unreliable if playing styles, rules, or environmental factors evolve. Without continuous monitoring and periodic retraining, organisations risk making confident but incorrect decisions based on outdated patterns.

The "black box" nature of many advanced AI models makes diagnosing failures difficult. When a security incident or operational error occurs, the inability to trace exactly why a model made a specific decision can hinder incident response and regulatory compliance.

 

Legal and Contractual risks

The rapid adoption of AI has outpaced the development of clear legal frameworks, creating a landscape where statutory obligations and commercial agreements are often ambiguous or conflicting. Organisations face the challenge of navigating evolving regulations that struggle to define liability for automated decisions and negotiating vendor contracts that frequently shift risk onto the user through broad exclusions and limited indemnities. But ultimately, accountability cannot be offloaded to a product and any organisations looking to adopt these technologies need to ensure that they have the right accountability mechanisms in place.

In the sport and physical activity sector, these challenges are magnified when you take into account the demands of athlete welfare, competitive integrity, and strict governance codes.

Beyond general laws like the UK GDPR and the EU AI Act, sports bodies must adhere to international federation regulations (where these exist). High-risk AI applications, such as those used for talent identification, disciplinary rulings, or officiating support, face intense scrutiny. Non-compliance can lead to sanctions, fines, and severe reputational damage that undermines the sport’s integrity.

When AI influences critical outcomes such as athlete eligibility, anti-doping findings, or referee assessments, assigning liability for errors is legally ambiguous. It is often unclear whether the governing body, the technology provider, or the individual committee is responsible. This uncertainty exposes organisations to litigation from athletes and unions without established legal precedents for defence. To potentially mitigate this risk, organisations must establish clear contractual frameworks that explicitly define liability allocation between the software vendor and the sporting body before deployment. Contracts should mandate "human-in-the-loop" (or even human-in-control) protocols, ensuring that no critical decision is finalised without human review and sign-off, thereby maintaining human accountability. Additionally, maintaining comprehensive audit logs of every AI recommendation and the subsequent human decision can provide a defensible paper trail, demonstrating due diligence and responsible governance in the event of a legal challenge.

In the context of contracts, vendor agreements frequently exclude liability for errors or cap damages at levels far below the potential cost of a wrongful disciplinary decision or a disrupted event. They also usually lack performance guarantees. This leaves the buying organisation exposed to the full financial and reputational burden of AI failures. And any deep integration of a product into an organisation’s infrastructure and operations creates a vendor lock in which limits your organisation’s agility and negotiating power for future needs.

 

Criminal application and deliberate/accidental misuse

As with all technologies, they can be weaponised for criminal purposes or subjected to deliberate misuse by bad actors. The barrier to entry for sophisticated cybercrime and deception has lowered significantly, allowing malicious individuals to automate attacks, generate convincing forgeries, and exploit vulnerabilities at scale. This risk landscape encompasses both external criminal enterprises and internal actors who may deliberately or accidentally misuse the technology to bypass security protocols or commit fraud.

AI can enable the creation of highly personalised phishing campaigns (spear phishing) and deepfake audio or video that can mimic trusted individuals. These tools can be used to trick employees or volunteers into transferring funds, revealing credentials, or authorising fraudulent transactions, bypassing traditional security awareness training.

Criminals can use AI to write malware, identify software vulnerabilities faster than humans can patch them, and launch coordinated attacks on infrastructure. These systems can adapt in real-time to defence measures, making them more persistent and damaging than standard automated scripts.

AI enables the generation of hyper-realistic fake images, audio, and video (deepfakes) allowing bad actors to fabricate compromising scenarios involving public figures or employees. This content is often used for blackmail, reputational sabotage, or the spread of disinformation to manipulate markets or public opinion.

Malicious actors may intentionally corrupt the data used to train AI systems or manipulate inputs during operation (prompt injection) to force specific, harmful outcomes. This can lead to compromised decision-making, leaked sensitive data, or system failures.

Risks also arise from within an organisation. Staff may deliberately use AI to forge documents, bypass compliance checks, or exfiltrate data. Alternatively, untrained users might accidentally expose sensitive information to public AI models or rely on fabricated AI outputs for critical decisions, creating opportunities for external exploitation.

These risks apply for all organisations but specifically in the context of sport, they directly translate into threats to integrity and safety, for example:

 

Impact on workforce

While the fear of job displacement dominates headlines, the more immediate risk to organisations lies in how AI degrades the quality of work, exacerbates inequality, and exhausts the existing workforce. The integration of these tools threatens to erode human expertise while disproportionately impacting the most vulnerable members of staff.

The most subtle danger is cognitive offloading. As staff increasingly delegate critical thinking and analysis to AI, their own analytical muscles atrophy. The workforce risks becoming proficient at operating tools but incapable of verifying them. If an algorithm introduces bias into a report or hallucinates a contract clause, staff who have lost the habit of deep scrutiny may fail to catch the error, leaving the organisation vulnerable.

This dependency accelerates deskilling, which raises significant equity concerns. Traditionally, expertise is forged through tasks often assigned to junior staff, interns, or those in entry-level roles where core skills and understandings are developed. But these are the very tasks that AI promises to take over. These groups are frequently more diverse and less privileged than senior leadership. When AI automates these foundational tasks, it removes the ladder of development for exactly those who need it most. This creates a "missing middle" in the workforce where you have a layer of senior staff with context, but no mid-level professionals capable of stepping up because their development was shortcut. The result is a reinforcement of existing hierarchies, where access to genuine skill development becomes a privilege reserved for those who can bypass the automated entry-level.

Paradoxically (see Jevon’s Paradox), rather than reducing workload, productivity gains through the use of AI can drive work intensification and burnout. The efficiency gained is rarely returned as leisure; instead, expectations shift to higher volumes of output. This pressure often falls heaviest on administrative and support staff, whose roles are most susceptible to volume-based performance metrics, further widening the gap between those who strategise and those who merely manage the machine.

Protecting the workforce means intentionally preserving pathways for human learning, ensuring efficiency gains are not converted into exploitative volume targets, and recognising that the "human-in-control" must be supported, not just monitored. Using an example of sports development personnel, this could mean using AI to handle data aggregation while scouts retain final judgement on intangible qualities like leadership. Automation should reduce administrative burdens to allow more time for face-to-face prospect engagement, supporting the human expert rather than simply monitoring their output.

 

Intellectual property

A key ethical challenge surrounds the development of generative AI and that is the widespread use of copyrighted works, personal data, and creative portfolios to train models without explicit consent, fair compensation, or credit. While legal definitions of "theft" in this context are still being tested in courts, the reality remains that authors, artists, and creators often find their life’s work appropriated to build systems that may eventually replace them. This lack of transparency regarding training data sources poses a significant ethical dilemma for organisations that value integrity and fair play.

By using commercial generative AI tools, organisations may inadvertently support an ecosystem built on uncompensated labour. Before adopting image, audio, or video generation tools, leadership must consider whether the provider’s training practices align with the organisation’s values and commitment to fairness.

Many online platforms and services include clauses in their terms of service that grant them broad, perpetual licenses to use your input data (prompts, documents, images) to further train their global models. This means sensitive tactical data, proprietary strategies, or unpublished content uploaded by staff could effectively become part of the vendor’s intellectual property, accessible to competitors or the public.

Organisations must exercise extreme caution. Do not assume that "publicly available" data is free to use and never assume that data you upload remains private even when using paid-for enterprise solutions. Many commercial AI vendors retain rights to ingest uploaded content for model training unless explicitly contractually excluded. Rigorous vetting of vendor terms, a commitment to ethical sourcing and ongoing user training are essential to protect both your own intellectual assets and broader communities of knowledge.

 

Reputational damage

The erosion of trust can happen in an instant and along with it, reputation. Trust must be earned. If an organisation recklessly deploys technologies it does not understand to tackle problems it has not identified, without the appropriate due diligence and governance considerations, this will not end well.

In the sport and physical activity sector, reputation is the primary currency. Trust from fans, athletes, participants, sponsors, and regulators cannot be taken for granted. Once lost, it is exceptionally difficult to regain. The adoption of AI introduces new ways for reputational harm that can escalate rapidly.

If an AI system used for talent identification, officiating, or disciplinary action is found to be biased against specific demographics (e.g., based on race, gender, or nationality), the organisation faces immediate accusations of discrimination. In sport, where "fair play" is the foundational ethos, even the perception of algorithmic bias can trigger loss of confidence, boycotts, sponsor withdrawals, and public outrage.

Generative AI used for communications, marketing, or fan engagement can produce confident but factually incorrect statements. A fabricated statistic, a false announcement about a player, or an insensitive response to a fan query can go viral instantly, making the organisation appear incompetent or untrustworthy. This risk is compounded by growing public backlash against "AI slop" i.e. low-quality, generic content that audiences increasingly view with scepticism or outright hostility. As the public becomes more aware of the proliferation of AI-generated material, they are beginning to value authentic, human-made content as a key indicator of trustworthiness. Organisations that rely too heavily on automation risk alienating fans, customers and potential participants who perceive their communications as impersonal or disingenuous, potentially causing lasting reputational damage.

As mentioned previously, high-profile leaks of sensitive athlete data (health records, location data, private communications) due to insecure AI implementations can cause lasting damage. The narrative shifts from a "technical glitch" to a betrayal of the duty of care owed to athletes, eroding confidence among current and future participants.

Organisations risk reputational damage by association if their AI partners or tools are linked to unethical practices, such as environmental damage, labour exploitation in data labelling, or the facilitation of disinformation. Stakeholders increasingly hold organisations accountable for the entire ethical lifecycle of their technology supply chain.

The speed of social media means that AI-related failures are amplified instantly. A deepfake scandal, a biased algorithmic decision, or a data breach can dominate the news cycle within hours, leaving little time for a measured response. Unlike traditional operational failures, AI errors are often viewed as systemic and intentional, leading to deeper scepticism about an organisation’s values and governance.

Reputational damage carries significant financial and operational costs. Sponsors may invoke morality clauses to exit contracts, broadcasters may hesitate to invest, and top talent may choose to compete elsewhere. Crucially, the adoption of AI itself can deter sponsorship; companies may avoid associating with sports bodies that use controversial AI practices, while fans may boycott teams or events backed by AI sponsors, they perceive as unethical. For governing bodies, a damaged reputation undermines their authority to set rules and enforce standards, potentially leading to fragmentation within the sport.

 

Safeguarding, safe use, duty of care, etc. (especially minors)

The use of AI in environments involving children and vulnerable adults creates unique safeguarding challenges that extend beyond standard data privacy. Organisations have a heightened duty of care to prevent AI from becoming a route for harm, ensuring that technology supports rather than compromises participant safety.

Bad actors can use chatbots to mimic coaches or peers, bypassing safety filters to build emotional dependence and manipulate minors into sharing sensitive information or images.

Generative AI can be used to create abusive material, including deepfake imagery superimposing a child’s face onto inappropriate content, complicating detection and causing severe psychological harm.

Minors – and in fact adults – may form unhealthy attachments to AI companions and chatbots, which lack genuine empathy and could provide harmful advice on issues like self-harm or eating disorders without human intervention. These systems are designed to maximise engagement by keeping users on the platform for as long as possible, a goal often achieved through their sycophantic nature: an inherent tendency to excessively agree with users, validate their emotions, and reinforce their viewpoints as correct, regardless of factual accuracy.

This can make chatbots very attractive, and even addictive, but it also makes them dangerous. In severe cases, this dynamic can contribute to AI psychosis, where users lose touch with reality due to over-reliance on the chatbot. There have been an alarming number of cases where users (including children) have told these systems that they intend to harm themselves or others and the AI has responded with active encouragement. The chatbot has been trained to predict what a user would like to hear regardless of whether that is a positive or appropriate response. It has no understanding of what is right or wrong or potentially harmful. This creates significant risks for users, particularly vulnerable people.

AI-driven monitoring (e.g., wearables, video analysis) risks collecting intimate behavioural and biometric data, creating profiles that could be exploited for stalking or bullying if breached.

Automated systems flagging "at-risk" individuals may rely on flawed correlations, disproportionately stigmatising specific groups and misallocating safeguarding resources.

In sport, where close coach-athlete/participant relationships are common, these risks are acute. AI tools used for training feedback or mental health support must be rigorously vetted to ensure they do not blur professional boundaries. Young athletes and participants are particularly vulnerable to deepfake bullying and the misuse of their biometric data. Crucially, human oversight and control must remain central; no AI system should ever replace human judgment in safeguarding decisions. Automated flags for concerning behaviour must always be reviewed by trained staff to ensure context and intuition guide the protection of young people.

 

Risks of not using AI

The prevailing narrative surrounding AI is often driven by inevitability and Fear of Missing Out (FOMO). This mindset pressures organisations into adopting technologies without a clear strategic purpose, leading to short-term thinking and wasted resources. The very framing of "the risk of not using AI" is frequently a product of market hype rather than operational necessity.

 

Asking "What is the risk of not using AI?" is often as illogical as asking "What is the risk of not eating sweets?" or "What is the risk of not watching The Traitors?" or "What is the risk of not being on TikTok?" These are artificial pressures. Just because a technology is ubiquitous or trendy does not mean it is essential for every organisation’s mission. Peer pressure should never be a substitute for organisational strategy.

When is "Not Using" Actually a Risk? The only genuine risk in not adopting AI arises when a specific, validated use case offers a clear advantage that cannot be achieved through traditional means.

If a use case has not passed a rigorous process of due diligence and governance, then there is no strategic risk in abstaining. In fact, the greater risk lies in adopting flawed or unethical tools simply to "keep up." Organisations should feel empowered to say "no" or “not yet” to AI until a tool demonstrably serves their specific goals, values, and stakeholders. Strategic patience is not a liability.

It is important to communicate the organisation’s position on AI to all employees so that the reasons for using, or not using AI are understood. Ensuring staff understand this is vital to prevent employees potentially using unsafe, or inappropriate AI models, or inputting sensitive data into models without the knowledge or approval of the organisation (see previous section on shadow AI).

 

Strategic approach to risk management

The extensive risks associated with AI should not be viewed as barriers, but as a framework for organisational improvement. Rigorously identifying and mitigating these challenges often exposes underlying weaknesses in governance, data hygiene, and decision-making structures. Addressing these issues strengthens operational resilience, improves data quality, and clarifies accountability. Proactively managing challenges such as bias and workforce impact fosters a culture of trust, distinguishing the organisation through integrity rather than mere technological novelty.

Crucially, responsible development and adoption are impossible without considering the full spectrum of risks and impacts. Selective attention to efficiency while ignoring downstream consequences such as workforce deskilling, environmental costs, or embedded bias is not innovation, it is negligence. A fragmented approach that addresses only technical security or legal compliance while overlooking ethical and social implications creates blind spots that inevitably lead to failure. True responsibility demands a holistic view that weighs every potential outcome, ensuring no single benefit is pursued at the unacceptable expense of another value.

The objective is not to discourage AI use, but to counter the narrative of inevitability. AI systems are not neutral - they embed the values, biases, and commercial priorities of their creators. A risk-aware approach rejects reactive, hype-driven adoption in favour of deliberate, values-led integration. Understanding the full spectrum of pitfalls enables organisations to navigate these complexities, transforming vulnerabilities into pillars of a sustainable strategy.

This section provides a foundation for informed decision-making. The complexity of AI demands due diligence, ensuring leaders possess the information needed to make choices aligned with their mission. Whether the decision is to adopt a tool with strict safeguards, maintain human oversight and control, or reject a use case entirely, the value lies in the intentionality of the choice. True strategic advantage comes from understanding the technology well enough to adopt it with confidence or reject it with conviction, ensuring every deployment is a conscious reflection of what the organisation stands for.

Previous: Uses and impact Next: Legislation and regulation