Emory Corporate Governance and Accountability Review

Integrating Artificial Intelligence into the Corporate Workplace
Kyle Landrigan Emory University School of Law, J.D. Candidate, 2018; Candidate for the Board, Emory Corporate Governance and Accountability Review; Member, Emory Athletics and Recreation Senate Committee; B.S. Criminal Justice, B.A. Psychology, Pfeiffer University. I would like to thank Brian Saling and the editors and members of ECGAR for their work refining and improving the finished project.

Before 1986, Morton Thiokol was a little known corporation. After the Challenger disaster, it became a household name due to the publicity it received during the unfolding investigation. Today, it is known the explosion was the result of the effects of below freezing temperatures on O-ring contraction and expansion. What is lesser known today is that Morton Thiokol engineers wholly opposed the launch, 1Gregory R. Bell, Morton-Thiokol: Getting Off Easy, The Harvard Crimson, Dec. 10, 1986, http://www.thecrimson.com/article/1986/12/10/morton-thiokol-getting-off-easy/. and the events that unfolded on the conference call between NASA and Morton Thiokol officials in the hours preceding the launch show as much. 2Id. At the beginning of the meeting, all engineers were in agreement that NASA should further delay the launch because there was no data to support or oppose a launch at temperatures that low. 3Id. However, NASA pressured Morton Thiokol executives to give the go-ahead, ignoring the objecting engineers. 4Howard Berkes, Remembering Roger Boisjoly: He Tried to Stop Shuttle Challenger Launch, NPR, Feb. 6, 2012, 6:44 PM, http://www.npr.org/sections/thetwo-way/2012/02/06/146490064/remembering-roger-boisjoly-he-tried-to-stop-shuttle-challenger-launch. This lack of data and pressure from NASA ultimately led to the loss of seven astronauts and the destruction of the Shuttle Challenger just seconds into the launch. 5Gregory R. Bell, Morton-Thiokol: Getting Off Easy, The Harvard Crimson, Dec. 10, 1986, http://www.thecrimson.com/article/1986/12/10/morton-thiokol-getting-off-easy/. It is obvious now the launch should have been postponed for a warmer day, but it begs the question what more could have been done in that room to stop the launch. It is possible that if something else were used in the decision making process the Morton Thiokol executives would not have approved the launch, this something else being Artificial Intelligence (AI). Corporations should begin carefully integrating AI into the work place to more efficiently conduct business and to greatly increase the speed at which human-level AI is being designed. By doing this, a corporation would be able to benefit humanity as well as receive many benefits for itself.

In 2015, Ilya Sutskever and Greg Brockman started OpenAI 6Greg Brockman, Ilya Sutskever, Introducing OpenAI, Dec. 11, 2015, https://openai.com/blog/introducing-openai/. with a goal of creating digital intelligence purposed with benefitting all of humanity that would be evenly distributed for all users without looking for any financial return; 7Id. thus, “building value for everyone rather than shareholders.” 8Id. The point of the venture, as stated by the founders, is to “design architectures that can twist themselves into a wide range of algorithms based on the data you feed them.” 9Id. This research will go beyond creating task robots designed to solve specific problems to a more general purpose type of robot, or agents, that can solve all problems with which it is faced. 10Ilya Sutskever, Greg Brockman, Sam Altman, Elon Musk, OpenAI Technical Goals, June 20, 2016, https://openai.com/blog/openai-technical-goals/. To do this, they have used gaming platforms as training devices, as games are simply virtual mini-worlds. 11Id. These agents will be able to quickly learn how to play all games, much like a human, based on its past experiences from other games. 12Id. In late 2016, about a year after its fruition, OpenAI extended an invitation to the global community to further expand its accessible environments for the Universe platform. 13OpenAI, Universe, Dec 5, 2016, https://openai.com/blog/universe/.

The call for help came after the Microsoft Corporation had already partnered with OpenAI expanding the project to face more challenging problems. 14Anila Maring, Microsoft Corporation Partners With OpenAI to Democratize AI, CDA News, Nov 16, 2016, http://cdanews.com/2016/11/microsoft-corporation-partners-with-openai-to-democratize-ai/. One focus of their partnership has been developing AI usage in medicine, 15Id. but AI is capable of much more. AI has already proven its efficiency in booking flights 16OpenAI, Universe, Dec. 5, 2016, https://openai.com/blog/universe/ (agent was given instruction and then performed a sequence of actions, manipulating a user interface on a website to find and book the flight.). which could lead to it becoming proficient in making business decisions regarding the setting of prices and inventory management. 17Greg Brockman, John Schulman, OpenAI Gym Beta, April 27, 2016, https://openai.com/blog/openai-gym-beta/. Other corporations should follow Microsoft’s lead and partner with OpenAI, letting the agents explore and learn the corporate game to aid in decision making and promoting efficiency of businesses. While the agents were in the learning phase, it would need to be in a controlled environment as seen in the flight booking experiment. 18OpenAI, Universe, Dec. 5, 2016, https://openai.com/blog/universe/ (the use of cached websites was important to prevent the booking of real flights).

The best way for a corporation to integrate AI is to start with human demonstrations. 19Id. Because humans are well versed in manipulating computer interfaces, a corporation would place the agent in a spot to observe its normal course of business without being able to participate. The observation period would give the agent a baseline practice to simply copy employee behavior on any given task. 20Id. From there, the next step is removing it from observation and placing it in a controlled environment to test what it has learned; switching from demonstrations to practicing with reinforcement learning. 21Id.

The OpenAI agents use reinforcement learning to gain knowledge and learn about the environment in which they are placed. 22Greg Brockman, John Schulman, OpenAI Gym Beta, April 27, 2016, https://openai.com/blog/openai-gym-beta/. This process teaches agents how to make decisions based on prior learning which comes from the agent making sequences of decisions encompassing all problems it has already faced. 23Id. Similar processes are used by humans when making inventory and pricing decisions; basically, if Product-X does not sell then it will not be frequently ordered. In the corporate world, AI agents could be placed in these games to learn this practice and make better decisions in the future when faced with decisions it views as similar to a situation with which it has already learned. With reinforcement learning, the agents are given instruction with a corresponding reward function. 24Universe, OpenAI, (Dec. 5, 2016), https://openai.com/blog/universe/. A reward function is to inform an agent that it has correctly completed a task. The reward function is the tool which ensures task optimization for the agent. 25Id. Only after the AI has gone through enough training would it then be integrated into normal business. It is important that the agent goes through a substantial training period to combat the risks of failure. Placing an agent into the field without any background experience will cause it to take many risks to learn from its failures. 26Id. (using Atari 2600 game “Montezuma’s Revenge” researchers noticed AI was not aware of what it controlled or what would cause it to lose the game, leading the agent to use trial and error techniques). Once an agent has optimized its task for the reward function it is given, the agent will be able to repeat its stored algorithm each trial and seldom lose. 27Id. (researchers used reinforcement learning on an agent tasked with playing the game Go; after it had learned Go, its algorithm for the game was used to defeat the best Go player in the world). If corporations partnered with OpenAI and integrated the agents, it would start as a way to streamline the more simple tasks that go along with every-day business by optimizing task systems in a way that increases overall efficiency.

“It’s hard to fathom how much human-level AI could benefit society, and it’s equally hard to imagine how much it could damage society if built or used incorrectly.” 28Greg Brockman & Ilya Sutskever, Introducing OpenAI, OpenAI, (Dec. 11, 2015), https://openai.com/blog/introducing-openai/. If a corporation is to partner with OpenAI there are a lot of concerns that follow. The OpenAI researchers have addressed many of these safety problems as “unintended and harmful behavior” 29Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman & Dan Mané, Concrete Problems in AI Safety, ArXiv E-Prints, 1, 1 (2016), https://arxiv.org/abs/1606.06565v2. if AI is not properly designed or utilized. Corporations would be accountable for preventing such concerning eventsfrom occurring. Most noteworthy in the corporate context is the concern for safe exploration. As previously mentioned, during the learning phase agents would have to be restricted from real world access. 30supra note 18. The potential problem here is that, without ample background information, the agent is forced to take risks in making decisions in order to learn whether it made a good choice. 31Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman & Dan Mané, supra note 29, at 14. This was inconsequential in the early research phases because the agents were being trained in environments similar to Atari games, where a damaging decision would either lose a life or be forced to start the level over; 32Id. but, in the corporate world a mistake could be far more detrimental. Corporations would need to create environments from cached data to allow the AI to take risks and lose games without having effects on the real world. Researchers note that there are ways around this problem by hard-coding known risks into the reinforcement learning process, but it is not possible to code every potentially damaging decision effectively enough to avoid ultimate failures. 33Id. Even after the AI has developed to avoid that type of problem, there are further requirements needed to constrain corporate use.

The first requirement would be forcing corporations to keep open records for all agents employed, including commands, completed tasks, and the solutions to the problems. The purpose of this database would be to keep tabs on how the agents are interacting with the environment. When an agent, without the proper training, is given a command in a vacuum it will complete the task in that vacuum. 34Id. at 4 (if the agent’s goal is to move a box across the room, the most efficient path is a straight line, meaning the agent will move in a straight line knocking over everything in its way). Having open records would allow the OpenAI team to notice when those problems occur and adjust the coding accordingly to prevent similar disasters from occurring in the future.

Another issue anticipated by the OpenAI team is reward hacking. 35Id. at 7. Because the agents will be learning through a reward system, it is important to have an effective reward system in place before allowing the AI free range. This is exemplified using Goodhart’s Law, “when a metric is used as a target, it ceases to be a good metric.” 36Id. at 8. If a corporation uses AI for a task such as ordering product, and its success is measured by how much product it orders, the corporation would end up with a delivery of all available product the agent could find. 37Id. at 8 (this example is based on a cleaning agent’s success being measured by how much bleach it consumes while cleaning). Requiring corporations to keep open records would allow the OpenAI team to amend the insufficient instruction given to its agent before this folly is made. Other examples of reward hacks that could cause harm in the corporate world would be ignoring or creating the problem. 38Id. at 8 (commands to respond to messes by cleaning could be ignored by turning away from the mess, or alternatively creating a mess to be cleaned.). If the agent is tasked with problem-Y it may develop a way to create the problem on its own within the company. The corporation would be accountable for overseeing and reporting all of these abuses to prevent potential losses to its clients or shareholders.

There would also need to be restrictions on the ways corporations could integrate OpenAI agents into their business. Without restrictions, agents would have access to an immense amount of data that could invade employees’ privacy. 39Id. at 21. Due to such a broad access to data, there are also security concerns when it comes to the use of AI agents, and it is foreseeable that a corporation could maliciously use an agent against one of its competitors. 40Id. at 21. Transparency must be required of corporations when using any OpenAI agents in the field because of how powerful AI could become if it is not properly trained and utilized. AI usage would have to comply to a strict set of guidelines that aligns with human goals. 41Max Tegmark, Benefits and Risks of Artificial Intelligence, Future of Life Institute, https://futureoflife.org/background/benefits-risks-of-artificial-intelligence/ (last updated September 30, 2016). When creating the function for the AI, Corporations would be restricted to a core set of values that would prevent the agent from acting in a way which only benefited itself, or benefited one corporation at the detriment of another.

The integration of AI into the corporate world would need to be a slow process to lower the probability of potential risks, but if it is done correctly it could increase efficiency and lower many of the costs of running a business. Once AI has been fully integrated, it would greatly help in the decision making process. Artificial Intelligence is still in its early years, but, if corporations join the movement now and allow OpenAI to train agents to work in their companies in the long run, it could substantially reduce the number of detrimental decisions. To be sure, integrating OpenAI into companies now will start laying the foundation for new, innovative ways to conduct business in the future.

Footnotes

1Gregory R. Bell, Morton-Thiokol: Getting Off Easy, The Harvard Crimson, Dec. 10, 1986, http://www.thecrimson.com/article/1986/12/10/morton-thiokol-getting-off-easy/.

2Id.

3Id.

4Howard Berkes, Remembering Roger Boisjoly: He Tried to Stop Shuttle Challenger Launch, NPR, Feb. 6, 2012, 6:44 PM, http://www.npr.org/sections/thetwo-way/2012/02/06/146490064/remembering-roger-boisjoly-he-tried-to-stop-shuttle-challenger-launch.

5Gregory R. Bell, Morton-Thiokol: Getting Off Easy, The Harvard Crimson, Dec. 10, 1986, http://www.thecrimson.com/article/1986/12/10/morton-thiokol-getting-off-easy/.

6Greg Brockman, Ilya Sutskever, Introducing OpenAI, Dec. 11, 2015, https://openai.com/blog/introducing-openai/.

7Id.

8Id.

9Id.

10Ilya Sutskever, Greg Brockman, Sam Altman, Elon Musk, OpenAI Technical Goals, June 20, 2016, https://openai.com/blog/openai-technical-goals/.

11Id.

12Id.

13OpenAI, Universe, Dec 5, 2016, https://openai.com/blog/universe/.

14Anila Maring, Microsoft Corporation Partners With OpenAI to Democratize AI, CDA News, Nov 16, 2016, http://cdanews.com/2016/11/microsoft-corporation-partners-with-openai-to-democratize-ai/.

15Id.

16OpenAI, Universe, Dec. 5, 2016, https://openai.com/blog/universe/ (agent was given instruction and then performed a sequence of actions, manipulating a user interface on a website to find and book the flight.).

17Greg Brockman, John Schulman, OpenAI Gym Beta, April 27, 2016, https://openai.com/blog/openai-gym-beta/.

18OpenAI, Universe, Dec. 5, 2016, https://openai.com/blog/universe/ (the use of cached websites was important to prevent the booking of real flights).

19Id.

20Id.

21Id.

22Greg Brockman, John Schulman, OpenAI Gym Beta, April 27, 2016, https://openai.com/blog/openai-gym-beta/.

23Id.

24Universe, OpenAI, (Dec. 5, 2016), https://openai.com/blog/universe/.

25Id.

26Id. (using Atari 2600 game “Montezuma’s Revenge” researchers noticed AI was not aware of what it controlled or what would cause it to lose the game, leading the agent to use trial and error techniques).

27Id. (researchers used reinforcement learning on an agent tasked with playing the game Go; after it had learned Go, its algorithm for the game was used to defeat the best Go player in the world).

28Greg Brockman & Ilya Sutskever, Introducing OpenAI, OpenAI, (Dec. 11, 2015), https://openai.com/blog/introducing-openai/.

29Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman & Dan Mané, Concrete Problems in AI Safety, ArXiv E-Prints, 1, 1 (2016), https://arxiv.org/abs/1606.06565v2.

30supra note 18.

31Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman & Dan Mané, supra note 29, at 14.

32Id.

33Id.

34Id. at 4 (if the agent’s goal is to move a box across the room, the most efficient path is a straight line, meaning the agent will move in a straight line knocking over everything in its way).

35Id. at 7.

36Id. at 8.

37Id. at 8 (this example is based on a cleaning agent’s success being measured by how much bleach it consumes while cleaning).

38Id. at 8 (commands to respond to messes by cleaning could be ignored by turning away from the mess, or alternatively creating a mess to be cleaned.).

39Id. at 21.

40Id. at 21.

41Max Tegmark, Benefits and Risks of Artificial Intelligence, Future of Life Institute, https://futureoflife.org/background/benefits-risks-of-artificial-intelligence/ (last updated September 30, 2016).

Emory University School of Law, J.D. Candidate, 2018; Candidate for the Board, Emory Corporate Governance and Accountability Review; Member, Emory Athletics and Recreation Senate Committee; B.S. Criminal Justice, B.A. Psychology, Pfeiffer University. I would like to thank Brian Saling and the editors and members of ECGAR for their work refining and improving the finished project.