AI and Berkeley Rent Control

By choice, I am very interested in Artificial Intelligence. Not by choice, I am very interested in rent control in Berkeley. Anxious to try out the emerging phenomenon that is AI, I asked ChatGPT to write an essay on the history of rent control in Berkeley. The result was published in the April issue of this newsletter.

The piece that ChatGPT produced was amazing in some respects and somewhat disappointing in others. First it was produced in a matter of seconds - not just quickly, but literally a matter of seconds. All the words were spelled right but spell-check has been around for decades. More impressive was the high level of readability. It had the quality of a well-written, college-level essay. It was hard to believe this was produced with absolutely no direct human input. It was written and edited by no one.

It was also fact-checked by no one. There were errors of omission and commission presented with equal authority to those facts it got right. Errors obvious to me with decades of exposure to the subject would have gone uncontested by someone new to the issue. I annotated my comments in italics as part of the piece in the April newsletter.

Curious, I gave ChatGPT another shot at the same task. I made the same request again. What I got was a surprise. All of the errors I found were gone. This does not mean they were corrected; they were mostly omitted or finessed. More importantly the bias was tempered with both sides of the issue presented. Here are some of the changes:

• Evictions were claimed to be a prime impetus for rent control in the first piece; they were not. Evictions were not mentioned in the newer version.
• The original piece incorrectly credited the City Council with enacting rent control in 1980. Enactment was in fact achieved by voter-passed initiatives in 1978 and 1980. The updated version avoids attribution and simply says that rent controls were enacted, albeit in 1979. Close enough.
• There is no repeat of the erroneous claim that rent increases were limited to eight percent.
• The original gave a largely-incorrect version of both the legal battles over the Berkeley ordinance and the passage and the impact of Costa-Hawkins. It actually had the City as the legal entity which passed C/H. Of course, C/H was a state law and the City of Berkeley was very strongly opposed to its passage. The new version makes no mention of the specific challenges or changes to the law, either judicial or legislative. It instead simply says that the ordinance has undergone several amendments and modifications.
• The original ChatGPT piece mentions a specific 3.5% allowable rent increase in effect today. This is not and was never the case. Initially rent increases were specified annually by the Rent Board; this was later amended when increases were linked to the CPI. This mistake is not repeated.
• Finally, the original piece had a clear bias in favor of rent control. It called it the cornerstone of the city’s housing policy, claimed it to be one of the most successful rent control programs in the country, and it stated that the ordinance continues to provide crucial protections for tenants and to preserve affordable housing. It concludes that Berkeley’s rent control program offers a model of how cities can protect and assure access to affordable housing. The newer version has paragraphs of equal length enumerating arguments attributable to both proponents and critics of rent control. The piece concludes that there is a continuing national discourse on the matter. The difference here is significant.

The newer piece is better in that it has no errors of significance and is far more balanced in its discussion of the effects of rent controls. It does so, however, by omitting relevant facts. The original got it wrong, but it did refer to the Fisher and Birkenfeld decisions and specifically mentioned Costa/Hawkins. These were absolute milestones in the history of Berkeley rent control. Their omission is a serious flaw in the new essay.

I am not sure how these changes occurred. I am not even sure I asked the same algorithm for an essay. I’d like to believe that whatever AI program produced the second version actually read my comments and made corrections accordingly. The whole idea of AI is that it has access to all the information, and it can and does process it. I am told, however, that [freebie] ChatGPT only incorporates available data through 2021. 

I appreciate that this editorial is more about AI than rent control. However, for rent control and any other matter of public debate, AI will be important, if not decisive, in determining public understanding and appreciation.  We fail to pay attention to the power of AI to our detriment.

For your own perusal, the two versions of The History of Berkeley Rent Control appear side by side on page 6.