Our previous article, The Importance of QA for Strong AI Guardrails in Games, talks about use cases for AI systems in video games, the importance of designing robust guardrails, and how to test them to mitigate faulty outputs.
The central challenge highlighted was the unpredictable nature inherent to AI systems – and the ease with which those systems can be exploited. These realities are not unique to video games.
There is a lot of propulsion driving AI development. Virtually all industries are integrating AI systems into their products and services – from tutoring platforms and smart home devices to enterprise tools and in-vehicle assistants.
But the risks carry over with the technology. While an NPC going off-script can dampen a player's experience, the consequences get considerably more serious when AI enters the real world.
In this article, we examine the risks AI can pose across tech-reliant industries, and why the lessons learned from implementing AI in gaming are more relevant than you might think.
Same Risks, Different Consequences
Experts understand the basics of how AI is built, training large models on vast datasets to recognize and predict patterns. However, the mechanism behind how a model arrives at a specific output is largely unknown – what the industry calls a “black box” – which makes finding a concrete solution quite challenging.
Regardless of the task it’s trained to do, AI comes with issues such as:
- Bias – discriminatory or unequal outputs rooted in training data
- Hallucination – incorrect information delivered with false confidence
- Safety – unsafe, harmful, offensive, or non-compliant AI behaviour
- Privacy & Data Leakage – unintended exposure of sensitive or confidential information
- Prompt Injection – manipulating the AI into overriding its instructions through malicious inputs
Just like players in a video game, people using AI-powered tools can deliberately (or inadvertently) find gaps in the system, leading to unforeseen outputs and possibly real-world consequences.
While the risks remain fundamentally the same, what changes across industries is the environment those failures occur in. The risk categories are consistent; the stakes are not.
What do these consequences look like in different fields?
Ed Tech
Intelligent tutoring systems, adaptive learning platforms, and AI writing assistants are among the many tools educational institutions are implementing to personalize and enhance learning at scale.
The challenge is that there is currently no reliable way to fully filter out flawed, outdated, or biased information from the datasets used to train these models. So, the model inherits these errors and biases directly, and they surface through its outputs.
This could lead to the dissemination of incorrect or misleading information, leaving students building on faulty foundations with misunderstandings going uncorrected. Biased training data can also lead to unfair treatment of certain student groups – quietly undermining the very experience these tools were designed to improve.
The fallout can leave institutions managing both the academic and reputational damages caused by a tool that wasn't adequately tested before deployment.
Enterprise AI
AI is being integrated into some of the most critical systems we rely on – from supply chain and logistics, predictive maintenance, and human resources to healthcare diagnostics and financial fraud detection, among others.
The same fundamental risks apply. Consequences scale with the environment.
What happens when a routine maintenance alert goes unflagged? What if an AI misinterprets a patient's medical file? Or a financial fraud detection system fails to catch what it was specifically designed to catch? The consequences are already playing out.
In 2023, a single incorrect answer from Google's AI chatbot wiped $100 billion from Alphabet's market value in a matter of hours.
In another case, a Samsung engineer accidentally leaked sensitive internal source code by uploading it to an AI tool for review. The incident prompted a company-wide ban on generative AI – a reminder that data shared with AI platforms can end up stored on external servers, outside the company's control.
A 2025 Nature Medicine study tested nine AI programs using 1,000 emergency room cases, keeping medical symptoms identical but changing patient details like race, gender, and income. The AI's recommendations changed based on those personal characteristics rather than the actual health condition.
Customer Support Chatbots
Customer-facing AI chatbots are now standard across industries, handling inquiries, processing requests, and representing brands at scale. As businesses race to deploy conversational AI across websites, apps, and service platforms, customer support chatbots have become one of the most visible – and vulnerable – points of contact between AI and the public.
A common exploit involves manipulating a chatbot's instructions to override its original purpose – effectively unlocking the full capabilities of the underlying AI model. With the right prompt, a customer service bot designed to answer product questions can be redirected to say almost anything, make commitments it has no authority to make, or bypass the restrictions it was built around. It's one of the most well-documented risks in deployed AI, and one of the most frequently overlooked.
In December 2023, a user manipulated a Chevrolet dealership's AI chatbot into agreeing to sell a $76,000 Tahoe for $1 – by simply instructing the bot to agree with anything the customer said. The car was never sold, but the chatbot was pulled from the site and the incident went viral with over 20 million views.
Proper guardrail testing would have caught this before launch. A QA professional thinking adversarially would have tested exactly what happens when a user attempts to override the bot's instructions, flagging the vulnerability before it became a public incident.
Automotive
For most AI-powered applications, there's a moment of review between a faulty output and its consequence. In automotive, that window is considerably shorter.
Navigation systems can misroute drivers leading to a missed turn, an unexpected detour, or a late arrival. Predictive maintenance tools that miss a developing fault, or flag one that isn't there, can mean unexpected breakdowns or unnecessary repair costs. And when a voice assistant misinterprets a command, it introduces another variable that requires the driver's attention while operating a vehicle in traffic.
The failures above are largely unintentional. But automotive AI also faces a more deliberate threat: prompt injection. In January 2026, researchers from UC Santa Cruz and Johns Hopkins demonstrated that self-driving cars would reliably follow instructions written onto road signs – including commands that could cause a vehicle to proceed through a crosswalk with a pedestrian present.
Smart Home & IoT
Smart home and IoT systems increasingly operate as interconnected, agentic workflows – meaning multiple AI systems communicating and handing off tasks to one another autonomously, in real time.
When everything functions as it should, users enjoy a whole new level of convenience, efficiency, security, and safety. However, the very sophistication that makes these systems agile also makes them fragile. The more agents in the chain, the more opportunities for error.
A misinterpreted voice command can trigger the wrong action across a connected system. A genuine security threat misclassified at one node may never escalate to the right response. And autonomous chains that execute without a human checkpoint can turn a simple misread input into an unintended sequence of actions.
Hardware, software and communication failures can occur randomly, and these systems are increasingly relied upon by people whose safety depends on them getting it right.
Tested in Games. Applicable Everywhere.
The risks outlined across these industries aren’t unsolvable. They just require Quality Assurance Testing by human professionals who can think adversarially, cultural context, and recognize when something is wrong even when it doesn't technically violate a rule. It’s the combination of people, processes and tools that offer a more complete solution for AI testing.
That capability doesn't change based on the industry. What changes is the domain knowledge needed to apply it effectively – understanding the specific users, environments, and failure modes that define risk in each context.
At Side, we've been providing QA testing on some of the most demanding technical projects in gaming for over 30 years. If you're navigating AI integration and want to talk about what thorough testing looks like for your product, get in touch with our team today.