Balancing AI Usage: The Importance of AI Guardrails

The Gist

Essential necessity. AI guardrails ensure responsible, safe, and ethical AI use.
Hallucination prevention. Guardrails mitigate AI hallucinations, reducing harmful outputs.
Marketing impact. Properly tailored guardrails enhance marketing, ensuring brand consistency and quality.

The term "guardrails" generally refers to mechanisms, policies and practices that are put in place to ensure that artificial intelligence (AI) systems operate safely, ethically and within certain predefined boundaries, particularly with generative AI models. These guardrails are important for ensuring responsible AI use and for minimizing risks and unintended consequences. For brands considering the use of generative AI, guardrails stand between them and a poor customer experience. This article will look at the reasons why guardrails are needed, the types of guardrails that are used and how the use of guardrails on generative AI impacts marketing.

Why Do We Need Guardrails?

Migüel Jetté, VP of AI R&D at speech-to-text transcription company Rev, told CMSWire that new AI tools are going public quickly with little to no safeguards, and some of them are years away from being properly trained. “As long as AI goes unregulated, companies will scramble to release unfinished or badly tested models — posing major problems for education, legal, and other industries,” said Jetté, emphasizing one of the reasons that guardrails play such a vital role at this point in the evolution of AI.

AI currently has a bad reputation, though it may be largely due to misconceptions based on science-fiction movies and now, the very vocal AI professionals calling for the regulation of AI, with some telling the public that AI will eventually cause humans to go extinct. Gaining the trust of the public is vital to the acceptance and adaption of AI across industries. “Setting guardrails for AI will improve the trust, adoption and potential of the tool. That greater accountability will help remove some of the fear around adopting AI, which when used and trained properly, has the opportunity for greatness,” said Jetté.

When OpenAI released ChatGPT-3.5 in November 2022, many people came forward to try it out and see what it could do. In February 2023, Microsoft announced that it was working with OpenAI to incorporate similar functionality into its Bing search engine. Within 48 hours of the announcement, over 1 million people signed up to be on the waiting list to try it out, and not long after, as testers began to use the generative AI model, strange results started showing up.

In one of the first such instances, a writer from The New York Times, Kevin Roose, tweeted that he had just had a conversation with Bing that left him feeling unsettled. In the conversation, Bing told Roose its real name (Sydney), described dark and violent fantasies, and tried to break up his marriage.

The same week that Microsoft announced the new Bing, Google demonstrated its own generative AI model named Bard as part of a live-streamed event in Paris. As part of the demo, Bard responded to a question with an incorrect response, and Google inadvertently used the incorrect response in the blog which introduced Bard to the world.

NewScientist rather immediately pointed out that Grant Tremblay at the Harvard–Smithsonian Center for Astrophysics had tweeted that the third statement in the demo was untrue — the first image was in fact taken by Chauvin in 2004 with the VLT/NACO using adaptive optics. The result of this press disaster resulted in the share price of Alphabet, Google’s parent company, dropping 9% in one day, wiping over $100 billion off its total value.

Similar mishaps occurred with ChatGPT and Bing, many of which were caused by what are called hallucinations — a phenomenon in which the AI model generates content that is not grounded in the input data or real-world knowledge. Essentially, the AI system "imagines" or "hallucinates" details, structures, or patterns that do not exist in the original training data or are not based in reality.

The data that an AI large language model is trained on accounts for the responses it creates, or in the case of Microsoft Tay, the engagements it has with users. Microsoft created Tay in 2016 to learn from interactions it had with 18-to-24-year-olds on Twitter. Soon after its release, it experienced a coordinated attack by groups of people who were determined to turn it into a Holocaust-denying racist. By the end of the week, Microsoft issued an apology, stating that it was "deeply sorry for the unintended offensive and hurtful tweets," and that it had taken Tay offline.

This type of behavior by AI is something that can be mitigated by guardrails, and this is the specific reason that they are used. Brands cannot afford to have these types of public mishaps, as they corrupt the trust that consumers have in them. AI can be used to effectively improve the customer experience, saving time, and increasing satisfaction — but only when they are properly aligned with a brand’s ethics and values.

CMSWire spoke with Vinod Iyengar, head of product at ThirdAI, a company making AI more accessible by training large language models (LLMs) on general CPUs (rather than GPUs). Iyengar is in a position to fully understand the need for guardrails, as his job revolves around the creation of LLMs for other businesses. Iyengar told CMSWire that essentially, we need guardrails for ethical, legal and reliability reasons.

Aside from the fact that unethical AI could generate offensive and harmful content, It could also come up with misleading, or even malicious information that incites hate and violence. More importantly, for AI to be accepted and adopted, Iyengar said that guardrails need to be in place to make sure AI is actually reliable. “We don’t want a system that makes up answers (known as ‘hallucinations’) or bring forth its bias,” said Iyengar. “Governments, tech companies, and society together need to come up with a common framework of AI guardrails to ensure we can safely trust and use the output of AI systems.”

Additionally, generative AI can incorporate inherent biases into its training data, and these biases must be eliminated before the AI model is used by a brand. “Bias is a big concern with AI, but with bias mitigation techniques that can detect and check a model for biases, we can build systems that monitor output and adjust as necessary to make it fairer,” explained Iyengar. “With all guardrails, the key is to advance AI training and fine-tune models with properly diverse and representative data.”

What Are Generative AI Hallucinations?

We’ve briefly mentioned situations when generative AI “hallucinates” when generating replies, but Bars Juhasz, military and defense A.I. researcher, founder and developer of Undetectable AI, a generative AI writing tool, provided CMSWire with an excellent analogy that explains the phenomenon.

“Imagine AI as a robot with a big brain that can talk and write,” said Juhasz. “Sometimes, this brain can be like a hyper kid making up wild stories; this is because it's ‘nondeterministic,’ which means it's unpredictable. By changing its mood settings ('temperature'), we can make it more sensible or creative. When it gets too creative and makes stuff up, it's called ‘hallucination.’"

Juhasz explained that researchers find these made-up bits by checking how excited and random the words are. “They also calm down the AI by asking the same question many times and averaging the answers, kind of like teamwork in solving a tricky math problem. This keeps the AI from getting too wild with its imagination.”

Juhasz went on to explain how guardrails work, stating that "LLM's, by their intrinsic nature, can be nondeterministic, which means its not going to output the same thing every time, and there’s randomness introduced through the temperature settings. Hallucinations are the result of this randomness, and there are multiple ways to both detect as well as deal with AI hallucinations." Juhasz emphasized that hallucinations can lead to brand disasters and deteriorate the customer experience when using generative AI for copy, content, or customer communications, so avoiding them is deeply important. "You want the generative text outputs to be factual and in line with the reference material you are providing it," he said.

What Kind of Guardrails Are There?

Generative AI models are capable of creating new content such as images, text, music and videos, and this ability can have specific risks and considerations. With adequate guardrails, brands can ensure that AI-generated content is safe and reliable. “With only a small amount of human feedback, we can train the AI to flag and avoid producing content that could be deemed as inappropriate, harmful, malicious, and so on,” said Iyengar.

Here are some of the guardrails that have been implemented for generative AI:

Content Filtering: Generative AI might create content that’s inappropriate, offensive, or biased. Using content filters to monitor and block such content is a must.
Plagiarism and Originality Checks: Generative models have a knack for mirroring training data, so it’s crucial to check for plagiarism and ensure originality.
Misinformation and Fact-Checking: Generative AI may sometimes unintentionally generate incorrect or misleading information. Fact-checking and correcting such information is extremely important.
Creative Attribution: This ensures that content that is generated by AI is properly attributed and not confused with human-created content, in order to respect copyrighted material and creative works.
Rate Limiting: This refers to implementing controls to limit the rate at which content is generated to sidestep spamming or manipulation.
User Feedback Mechanism: Brands should always create feedback channels for users to chime in on content issues, especially biases, misinformation or inappropriate material.
Controlled Generation: This allows brands to keep AI on a creative leash, guiding its output within certain themes, styles or borders.
Impact Assessment: Evaluating the potential societal impact of generated content is very important, especially in cases where the content might have wide reach or influence public opinion (similar to Facebook and Twitter impacting election results in 2020).
Deepfake Detection: In cases where generative AI is used to create realistic images or videos (aka deepfakes), this refers to using mechanisms to detect and label such content to prevent misinformation and manipulation.

Brands should keep in mind that guardrails are not unbreakable, and in fact, there are many people who enjoy trying to “jailbreak” them. Regina Grogan, an artificial intelligence researcher, spoke with CMSWire about the use of guardrails. She said that generative image models typically have guardrails to prevent users from creating pornographic or inappropriate content. “However, if a person tries just a little bit, they can easily break them. Within one or two prompts, I can break these guardrails,” said Grogan.” She suggested that guardrails are still put in place to prevent average users from creating such content.

Grogan said that large AI brands, such as Google, Microsoft and OpenAI, use guardrails to protect themselves from lawsuits, which is also a reason for other brands to use them. Although it’s not as likely for a customer to use Brand B’s AI chatbot to create explicit or dark content, people have been trying to break the guardrails that have been put in place since the generative AI models were released. As such, the use of guardrails is an iterative process, where guardrails are created, implemented, and tested by users, and then updated as needed to remain effective.

How Do Guardrails Impact Marketing?

Using guardrails in marketing with generative AI is very similar to using them for other reasons. For marketers, it is also about setting rules and guidelines to ensure that the content that is created by AI stays in line with the brand's values, follows rules and regulations, and meets ethical standards.

There are many areas of marketing that are or can be impacted by the use of guardrails, including brand consistency, content quality, regulatory compliance, the prevention of the aforementioned hallucinations, data privacy, crisis prevention, context awareness and the overall customer experience.

“For marketing, it’s important to ensure that the output of AI is actually factually correct, adheres to your brand guidelines, speaks in your voice and overall is relevant to your messaging,” said Iyengar. Ideally, brands that plan on using generative AI as part of their business should implement an AI model that is trained on their own private corpus so it learns from their original content.

Additionally, specific guardrails are going to be useful for one brand, but not for another. Each brand must determine how its goals would be impacted by the use of specific guardrails. “Online retailers for example, have been fighting the good fight against scam reviews. Amazon just came out saying it is using AI exactly for this cause. For retailers, they’d want AI guardrails that focus on content or review filtering,” said Iyengar.

Different industries would also find certain guardrails more applicable than others. “Consider an insurance brand that’s required to be fair and impartial when writing policies. In this case, the brand would need a system with bias mitigation guardrails, to ensure the company doesn’t inadvertently refuse a policy based on a person’s race, age, sexual orientation, etc.,” said Iyengar.

How Much Is Too Much?

Both ChatGPT and Bing, and Bard to a degree, have been criticized because many users feel that the guardrails that were placed upon them are too restrictive, rendering them useless, leading many to question, how much is too much?

Learning Opportunities

Webinar

Sep

CMS Briefing: A Live Look at What’s Next in AI-Driven Platforms

Learn how leading organizations are using AI‑driven tools to publish faster, personalize smarter and stay secure.

Webinar

Sep

Call Spoofing Trends Financial Institutions Can't Afford to Ignore

Don't let robocalls sabotage your customer relationships. Learn how to secure your voice channel.

Webinar

Oct

Agentic AI Playbook: Real-World Customer Service Use Cases You Can Deploy Now

Boost self-service by 30% and slash call volume by 63% with agentic AI.

Webinar

Oct

Beyond Storage: Smarter Content, Bigger Impact with DAM + AI

Discover how the DAM + AI duo makes content smarter, stronger and more accessible.

Webinar

On demand

Ready or Not: How Data-First Organizations Are Unlocking Agentforce Potential

Learn how to cut through the noise, activate Agentforce and build a Salesforce AI strategy that actually delivers.

Watch Now

Webinar

On demand

CX Reality Check: What Your Customers Are Really Thinking

Watch Now

Webinar

Sep

CMS Briefing: A Live Look at What’s Next in AI-Driven Platforms

Learn how leading organizations are using AI‑driven tools to publish faster, personalize smarter and stay secure.

Webinar

Sep

Call Spoofing Trends Financial Institutions Can't Afford to Ignore

Don't let robocalls sabotage your customer relationships. Learn how to secure your voice channel.

Webinar

Oct

Agentic AI Playbook: Real-World Customer Service Use Cases You Can Deploy Now

Boost self-service by 30% and slash call volume by 63% with agentic AI.

When asked just that question, Iyengar suggested that it really depends on the use cases and the industry. “Healthcare and finance guardrails would need to be vastly different. It depends on who the audience or user of the product is going to be. A ChatGPT or Bard that is open for everyone on the internet will likely need to err on the side of caution,” said Iyengar. “An AI that might be given to students or kids will need even more guardrails."

There are instances where it may be beneficial to not use many guardrails or even none at all. “An AI that is going to be used by a limited number of users for specific purposes may need other guardrails but could be more open as far as content generation is concerned,” said Iyengar. “In certain other domains, there could be a case made for actually letting a model go wild. For instance, researchers in pharma or drug development might be interested in letting a model ‘hallucinate’ pretty heavily to encourage the discovery of new protein sequences that could lead to new research,” said Iyengar.

Final Thoughts on AI Guardrails

AI guardrails are essential in marketing to safeguard brand integrity and customer trust by mitigating risks like content inaccuracies and biases. Brands must adopt and regularly update tailored guardrails to strike a balance between restriction and flexibility, ensuring AI effectively enhances the customer experience without compromising ethics and trust.