PDF Summary:Generative AI, by Harvard Business Review

Generative AI is transforming how businesses operate, but many organizations struggle to understand how to implement it effectively while managing the associated risks. In Generative AI, Harvard Business Review explains how these systems work, what they require to function, and how companies can use them responsibly.

You'll learn how generative AI can enhance creativity, streamline sales processes, and improve innovation across teams. The guide also covers practical implementation considerations, including data infrastructure requirements, integration strategies, and the importance of problem formulation. Harvard Business Review addresses the ethical frameworks and safeguards companies need to deploy AI responsibly, from managing bias in datasets to protecting data privacy and navigating emerging legal risks around intellectual property.

(continued)...

Integrating internal, public, and third-party data can enhance the AI's capability to comprehend context, forecast queries, and execute commands. However, it's important to verify the accuracy of external data against your internal information before incorporating it into the underlying dataset.

How Generative AI Writes Code

Generative AI can write code that connects your different data sources into one repository by encoding the database-related tokens into a high-dimensional space and then decoding them into syntactically valid code. This process allows the model to generate code that a database engine can execute. The authors explain that the model represents each token as a vector in a high-dimensional space, capturing the relationships between different tokens. By applying the same transformation repeatedly, the model can predict the next token in a sequence, allowing it to generate code that follows the statistical patterns of the programming languages it was trained on. This approach enables the model to produce code that is not only syntactically correct but also functionally coherent, making it a valuable tool for automating complex data integration tasks.

Implementation & Scaling

Merging generative AI with existing systems can provide quick value. According to the authors, it can enhance the applications salespeople utilize to draft emails or create sales presentations and proposals. It can also improve the quality of AI-suggested content by integrating customer sentiment insights. These upgrades can occur in the background, allowing users to benefit without having to relearn how to use the app's features.

Buying is faster than building when implementing solutions. Although developing a bespoke AI-driven system is more flexible, it's a process that demands significant time and resources. Purchasing a pre-existing application minimizes the demand for specialized internal talent and simplifies staying updated with rapidly evolving technology.

When to Buy and When to Build

When deciding whether to buy or build, consider whether the generative AI application needs to encode your organization’s unique decision logic. If it does, building a bespoke system may be necessary. For example, a financial services firm might need a generative AI system that can generate investment recommendations based on its proprietary risk assessment models and client profiling techniques. In this case, a pre-existing application may not be able to capture the firm’s unique approach to risk management and client service. On the other hand, if the generative AI application doesn’t need to encode unique decision logic, buying a pre-existing application may be a more efficient and cost-effective option.

Strategic Safeguards, Risks & Ethical Frameworks

Organizations must ensure the ethical, responsible, and transparent application of generative artificial intelligence. The authors argue that this tech ought to enhance human abilities and strengthen communities, rather than take over their functions. Organizations have an important responsibility to implement generative AI so that it improves—rather than reduces—the workplace experience for their staff and clients. This involves maintaining accuracy, safety, honesty, support, and eco-friendliness; reducing risks; and removing biased outcomes. The dedication should go beyond short-term business interests, covering wider societal duties and ethical standards for AI.

(Shortform note: In The Age of Em, economist Robin Hanson explores a future where advanced digital minds, or “ems,” take over many human functions. He argues that this scenario could be acceptable or even desirable, as ems could vastly outnumber humans, rapidly grow the economy, and largely replace humans in practical roles. In this world, biological humans would mostly live off the returns to their assets, need to do little work themselves, and could still have comfortable lives alongside a bustling, em-run civilization.)

Generative AI solutions sometimes can't grasp business or emotional contexts or recognize when they make errors or cause harm. Human involvement is necessary to assess the outputs for precision, identify bias, and verify that models function as designed. Generative AI requires ongoing supervision and can't work in a way where they're left to function without oversight. Organizations might begin by exploring methods to automate how they review by gathering data about artificial intelligence and creating typical strategies to mitigate particular risks. They might also think about funding ethical AI education for managers and front-line engineers to equip them for evaluating AI tools. If resources are limited, they can focus on testing the models most likely to cause harm.

Automating Review With Model Cards

One way to automate review while still supervising generative AI and focusing on the models most likely to cause harm is to require every generative AI system to have a standardized “model card” that describes its intended uses, data, and limitations. These model cards could be fed into a central registry that automatically flags the systems with the most sensitive impacts for more frequent human testing and review. This approach would help organizations systematically understand and compare models across different use cases, as the authors suggest, while also directing limited testing resources toward the areas of greatest potential risk.

Businesses ought to create AI protocols with an ethical structure to reduce risks and maximize benefits. Harvard Business Review notes that these practices guarantee AI's responsible, safe, and honest use. They also ensure AI is empowering and sustainable. In addition, these practices help companies avoid bias and fulfill their wider obligations to society. Companies can create a strategy for AI usage and ensure their AI objectives align with their business goals, thereby developing ethical practices for AI.

(Shortform note: The authors’ suggestion to align AI objectives with business goals to create AI protocols with an ethical structure may not work for all companies. In The Age of Surveillance Capitalism, Shoshana Zuboff argues that some companies’ business goals are inherently at odds with the interests of their users. She explains that surveillance capitalism unilaterally claims human experience as free raw material for translation into behavioral data, and it is a parasitic economic logic in which our lives are subordinated to mechanisms of extraction, prediction, and behavior modification that operate in disregard of our rights and the future of democracy.)

They can also offer various ways for employees to voice concerns, like a confidential phone line, email list, designated channels on social platforms like Slack, or groups that focus on particular issues. Incentivizing staff to voice concerns can also work well. Some companies have created ethics advisory groups—made up of employees across departments, outside specialists, or a combination—to provide input on developing AI. Lastly, maintaining open communication channels with stakeholders in the community is essential to preventing unintentional outcomes.

The Risks of Backfiring

A major risk of these strategies is that they can backfire if employees don’t feel safe or protected when voicing concerns. If employees feel that their concerns aren’t taken seriously or that they might face retaliation, they may be less likely to speak up about serious issues. Similarly, if ethics advisory groups are seen as merely symbolic or lacking real influence, they can undermine trust in the process. To avoid these pitfalls, companies must ensure that reporting channels are truly confidential, that there are clear protections against retaliation, and that there’s visible follow-through on concerns raised.

Governments and organizations are working to regulate artificial intelligence and establish ethical guidelines. The authors note that the US has passed 21 AI-related bills into law, including rules governing the use of facial recognition technology in legal proceedings and the establishment of an AI Division tasked with examining all AI applications by the state government and recommending an ethical code for AI in the state. The EU is contemplating a law featuring a framework to categorize the risks artificial intelligence might present to individuals' health, safety, or basic rights.

(Shortform note: Since the publication of this book, the EU has adopted the Artificial Intelligence Act, which means that the law is now binding and enforceable. This means that the EU has moved beyond merely contemplating the law and has taken concrete steps to implement it. The adoption of the law marks a significant milestone in the EU's efforts to regulate artificial intelligence and ensure that it is developed and used in a way that is safe, ethical, and respects fundamental rights.)

The African Union created a committee focused on AI, and the African Commission on Human and Peoples' Rights endorsed a resolution concerning the human rights implications of AI, robotics, and other emerging technologies in Africa. In 2021, China enacted a law on data protection, setting up guidelines for obtaining users' consent to collect data, and has also introduced a distinctive policy to regulate "deep synthesis technologies" used in deepfakes. The UK government announced a strategy to use current regulations as guidelines for novel AI tech.

(Shortform note: These initiatives are important because they show how AI rules are starting to fragment globally. For example, China’s data protection law and the EU’s GDPR have different requirements for how companies can move data across borders. This means companies might have to store data in specific countries or follow different rules depending on where they operate. A research article found that while many countries agree on basic AI principles like transparency and fairness, they often disagree on how to put these ideas into practice.)

We'll now discuss building robust safeguards and addressing emerging risks.

Building Robust Safeguards

To build robust safeguards for generative artificial intelligence, you must ensure datasets are representative and free from harmful biases. Harvard Business Review emphasizes that if a system is so big it's incomprehensible, it can't be completely purged of potential biases. Developers need to comprehend and record the dataset's inherent risks, which might require a smaller one.

Moreover, teams that are varied, particularly those that involve members from marginalized groups, can enhance the odds of having diverse identities and viewpoints represented in the data collected and generated for training models. This approach also aids in pinpointing previously unacknowledged biases or blind spots in the data.

The Limits of Technical Solutions to Bias

While these safeguards are important, they can also create new problems. In Data Feminism, Catherine D’Ignazio and Lauren F. Klein argue that technical solutions to bias, like representative datasets and diverse teams, can give a false sense of progress. They argue that these approaches often treat injustice as a problem to be solved by better data or by adding a few underrepresented people to a project. This can lead organizations to believe they've addressed bias when they've only made superficial changes. The authors warn that this mindset can actually reinforce existing power structures by making it seem like meaningful change has already happened.

Another important safeguard is to protect data confidentiality and safety. The authors explain that systems using private employee and customer data can be susceptible to malicious parties. A minor flaw might expose an organization to breaches. Blockchain advancements may assist. Blockchain technology provides a secure, shared database that logs data transactions. It’s presently applied in areas such as developing payment platforms and cryptocurrencies.

To safeguard data privacy and security, gather information about how your AI was created, and then assess if it's suitable for handling secure data. Maintain current technology systems and allocate budget for securing the software.

Blockchain and Data Privacy

While blockchain technology can help protect private employee and customer data, it can also undermine confidentiality. Michèle Finck, a law professor, explains that blockchain’s core feature of immutability sits in structural tension with the General Data Protection Regulation, because storing personal data on an append-only distributed ledger can make it practically impossible to fully comply with data subject rights such as erasure and rectification, and she argues that this incompatibility is particularly acute where identifiable or easily re-identifiable information is written on-chain rather than kept off-chain under a conventional data controller who can modify or delete it in response to legal obligations.

Addressing Emerging Risks

Businesses must address the legal risks of employing generative AI. Harvard Business Review points out that these include intellectual property infringement and unresolved legal questions, such as whether AI-generated creations violate copyrights, patents, or trademarks and the ownership of the output generative AI platforms produce for you or your customers. Organizations must grasp these risks and how to safeguard themselves before they can take advantage of AI tools.

(Shortform note: AI law experts argue that boards of directors have a fiduciary duty to oversee their companies’ legal exposure to AI-related risks. This means that decision-makers must understand these risks before they can authorize significant AI deployments. Otherwise, they could be held personally liable for failing to protect the company from foreseeable legal problems.)

Businesses should assess their transaction agreements to include protective measures in contracts. They need to request that generative AI platforms' terms of service ensure they have appropriate licenses for the data used in training. They should also require comprehensive indemnification against possible intellectual property violations due to a lack of adequate data input licensing by the AI companies, or because the AI's self-reported outputs aren't flagged for potential violations. When any side employs generative AI, companies need to include disclosures in their contracts with vendors and customers. Additionally, they should outline how both parties will aid in registering who authored and owns the works.

(Shortform note: To ensure that your company is protected from generative-AI-related risks, consider implementing a mandatory generative-AI contract addendum. This addendum should be attached to every agreement involving generative-AI use, regardless of whether your company or the other party is using the technology. The addendum should be regularly updated by your legal and risk management teams to reflect the latest developments in generative-AI technology and regulations. By making this addendum a standard part of your contracting process, you can ensure that all parties are aware of their obligations and that your company is consistently protected across all transactions.)

Vendor and customer agreements may contain AI-specific terms appended to confidentiality clauses to prohibit the party receiving the information from entering the disclosing party's confidential information as AI prompts. To minimize unforeseen usage risks, some prominent companies have developed generative AI checklists for their clients to use when modifying contracts, evaluating the implications of each clause related to AI. Companies employing generative AI or partnering with vendors who do should ensure their legal advisors are informed about the extent and manner of that usage, as legal regulations will keep changing swiftly.

Maintaining a Centralized Register of AI Systems

The National Institute of Standards and Technology (NIST) recommends that organizations maintain a centralized register of all AI systems, including those used by vendors and customers. This register should categorize each system by use case and risk level, allowing for periodic review of high-risk systems as regulations evolve. By labeling every vendor and customer agreement that involves generative AI, companies can ensure that their legal teams are aware of all AI-related provisions and can proactively update them as laws change. This approach also helps identify which agreements require more frequent legal review based on the risk profile of the AI use case.