Picture walking into city hall on a Monday morning. Printers hum. Clerks call ticket numbers. Residents form a line, waiting to ask questions about benefits, permits, and local programs. Behind the counter, a caseworker checks a dashboard that ranks files by urgency. On another screen, a traffic model predicts an afternoon bottleneck and proposes a timing plan for three intersections. The visible work looks ordinary. The invisible layer is a set of models that scan data, estimate risks, and surface recommendations, allowing people to act sooner and with clearer context.

AI in public policy weaves these tools throughout the full policy cycle. During agenda setting, text analysis helps sift through public comments and case notes to spot patterns that matter. In policy design, simulation explores what might happen under different rules before anything is put in place. In service delivery, triage models sort heavy caseloads so staff spend time where their judgment is needed most. In oversight, monitoring looks for drift, errors, and disparities, then alerts auditors when patterns change. The aim is a workflow that is faster to diagnose problems, clearer to explain, and easier to audit.¹

Trust grows when guardrails are visible. A risk approach asks three simple questions. What could go wrong? How likely is it? Which controls reduce risk to a level the institution can accept? The NIST AI Risk Management Framework turns those questions into everyday practice. Map the context and people affected. Measure the model and its harms. Manage with controls like documentation, testing, and human oversight. Monitor performance after launch and keep records that explain decisions. These steps give agencies a repeatable way to judge tools, compare vendors, and decide when to scale or stop.¹

Rules matter in daily operations. Canada’s Directive on Automated Decision-Making ties the use of AI to a required Algorithmic Impact Assessment and a risk tier that sets safeguards. Higher-impact systems need human-in-the-loop escalation, public notice to affected individuals, an explanation that a layperson can read, and clear routes to challenge an outcome. The policy also expects reproducible records, model documentation, and channels for redress, which keep accountability visible to program managers and auditors.²

Evidence demonstrates why the blend of analytics and oversight leads to better outcomes. Researchers studying pretrial decisions in New York City trained predictive models and ran policy simulations on detention and crime outcomes. At the same jailing rate as judges, simulated choices reduced crime meaningfully. Holding crime constant, detention could drop while maintaining public safety. The lesson is to reveal feasible operating points and tradeoffs that were hard to see before, then place those options inside a supervised process with clear pathways for review and appeal.³

Scale shapes how this lands. A national agency might run mature data pipelines, model risk teams, and standardized model cards. A small municipality might start with focused tools for permit triage, traffic timing, graffiti dispatch, or missed-pickup prediction. Both can work from the same playbook. Define the decision and the risk. Document the data and assumptions. Test for performance and disparate impact. Assign a human reviewer with the authority to pause or reverse an outcome. Publish what can be disclosed and invite feedback.^1,2

When these elements line up, AI becomes part of ordinary public administration. Staff get earlier signals. Residents receive clearer notices. Leaders see tradeoffs and outcomes in a form they can debate. The institution learns in shorter cycles and keeps a record of how and why it made choices. The goal is a public sector that is responsive, lawful, and worthy of trust.

Risk Management Framework: A structured method for governing models across their life cycle. It organizes work into mapping context and stakeholders, measuring performance and harms, managing with controls like documentation and human oversight, and monitoring after launch so agencies can compare tools, set thresholds, and decide when to scale or stop.

Algorithmic Impact Assessment: A pre-deployment review that rates system impact and ties risk tiers to safeguards. It drives requirements such as human escalation, plain-language explanations, public notice to affected people, reproducible records, and clear routes to challenge outcomes.

Human-in-the-Loop: A design choice that ensures trained staff can review, override, or escalate automated results. It assigns real authority to caseworkers or reviewers, makes accountability visible, and preserves a path to redress in higher-impact uses.

Transparency: Practices that make systems understandable to non-experts. Agencies publish inventories and documentation, provide their reasoning in plain language, use formats like model cards or counterfactuals where appropriate, and maintain audit trails so decisions can be examined and appealed.

Disparate Impact: A measurable difference in outcomes across protected groups that can result from a policy or model without explicit discriminatory intent. Teams assess it by comparing selection rates, error rates, or calibration across groups, then use those findings to adjust design, safeguards, and oversight.

Artificial intelligence first entered public life as an idea before it became a tool. In 1956, the Dartmouth workshop set the tone for AI as a field. The seed was the 1955 proposal by computer scientists John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, which called for a summer study of language, abstraction, and self-improvement in machines.⁵ The vision eventually touched government work, as policy always follows the progress of information tools.

Proof that programs could learn arrived soon after. In 1959, computer expert Arthur Samuel published a study of a checkers program that improved with experience.⁶ The paper showed two pathways for machine learning and hinted at something public agencies would later value: systems that adapt as new data arrives.

For decades, AI matured in labs and industry while governments focused on data collection, statistical reporting, and early e-government. A turning point in policy thinking came in 2016, when the White House Office of Science and Technology Policy released Preparing for the Future of Artificial Intelligence. The report mapped opportunities and risks, set out research priorities, and encouraged agencies to build capacity for responsible AI use.⁷ It marked the moment when AI governance moved from scattered memos to a coherent federal agenda.

The same year, a wave of public-facing scholarship made the stakes vivid. Data scientist Cathy O’Neil’s Weapons of Math Destruction told stories of scoring systems that could entrench inequity in hiring, education, and credit, offering language that policymakers could use in hearings and rulemaking.⁴ Two years later, Scholar and journalist Virginia Eubanks’ Automating Inequality documented how welfare, housing, and child-protection tools shaped life for low-income families, urging governments to center dignity and due process when they adopt automated systems.⁸ These books gave officials a human lens on technical choices.

Also in 2018, Computer scientists Joy Buolamwini and Timnit Gebru published “Gender Shades,” an evaluation of commercial facial-analysis systems that showed large accuracy gaps across gender and skin-tone subgroups.⁹ The study provided clear methods, public benchmarks, and numerical error rates that regulators and procurement teams could understand. It accelerated audits and policy conversations about fairness, transparency, and accountability in high-stakes public uses.

International coordination followed. In 2019, the OECD adopted the first intergovernmental principles on AI, centering human rights, transparency, robustness, and accountability. The recommendation gave governments a shared vocabulary for trustworthy AI and encouraged national strategies that paired innovation with safeguards.¹⁰ The document became a reference point for many public-sector roadmaps and guidance notes.

Standards work deepened that foundation. In 2023, the U.S. National Institute of Standards and Technology released the AI Risk Management Framework, a practical playbook that asks agencies to map context, measure performance and harms, manage with controls, and monitor over time. The RMF’s structure fit government workflows and procurement, making it easier to translate principles into daily practice.¹

Lawmakers advanced binding rules in parallel. The European Union finalized the AI Act in June 2024, a risk-based law with obligations for high-risk systems and transparency duties for specific uses. It set timelines for application and enforcement and pointed to future technical standards that will shape public-sector deployments across the bloc.¹¹ The regulation established a legal floor that many other jurisdictions began to study as they drafted their own approaches.

By the mid-2020s, AI in public policy had a recognizable shape. Researchers had framed the promise and the pitfalls. Authors and advocates had supplied real cases and language that the public could grasp. Standards bodies had offered a method for risk, documentation, and continuous monitoring. Legislatures had moved from guidelines to law. From McCarthy and Samuel to O’Neil, Eubanks, Buolamwini, and Gebru, the story traces a path from early experiments to systems that require rights-aware design, careful audits, and clear routes for challenge and appeal.

John McCarthy

A computer scientist who helped launch the field, McCarthy coined the term “artificial intelligence” and organized the Dartmouth workshop in 1956. That meeting set research on a path that later shaped how governments think about automated reasoning, public data, and decision tools.

Arthur L. Samuel

A pioneer of machine learning, Samuel built a self-improving checkers program and published his landmark paper in 1959. His work showed how models could adapt as new information arrived, laying early groundwork for public systems that learn from operational data.

Cathy O’Neil

A data scientist and author, O’Neil published Weapons of Math Destruction in 2016. The book brought wide attention to harmful scoring systems in hiring, education, and credit, giving policymakers a clear vocabulary for fairness, transparency, and accountability in automated decisions.

Virginia Eubanks

A scholar and journalist, Eubanks released Automating Inequality in 2018. Through case studies of welfare, housing, and child protection programs, she urged agencies to protect due process, explain outcomes in plain language, and offer accessible routes to challenge decisions.

Joy Buolamwini

A computer scientist and founder of the Algorithmic Justice League, Buolamwini coauthored “Gender Shades” in 2018. The study documented large accuracy gaps in commercial facial analysis across gender and skin tone, prompting audits, documentation practices, and tighter procurement standards in public applications.

AI in public policy helps agencies uncover how information flows shape outcomes across programs. It turns raw data into signals that guide staffing, budgets, and enforcement with greater consistency. The approach reaches transportation, health, benefits, housing, and inspections. By testing tools before full rollout and measuring real results, teams minimize delays, target resources, and make decisions they can explain to the public.

Green waves, calmer commutes

Morning traffic tells the story of a city. Adaptive signal control listens to that story in real time and edits the next chapter, one intersection at a time. When sensors and controllers coordinate green lights across a corridor, cars move in platoons, buses hold schedules, and side streets clear without a puzzle of missed cycles. Federal highway evaluations of adaptive systems reported measurable cuts in travel time and delay, along with fewer stops during peak travel hours.¹² Agencies like these programs because the wins stack up in small increments that commuters feel right away. Technicians get dashboards that show which signals are slipping. Planners get before-and-after evidence they can bring to council meetings. The city gets a tool that keeps working while crews focus on bigger fixes.

Deployments have put numbers on the board. Evaluations of SCOOT software in the UK reported a drop in intersection delay compared with fixed plans. Oakland County’s SCATS conversions saw travel time improvements at different times of day. Turin’s UTOPIA delivered boosts to average bus speeds while keeping other movements flowing.¹³ These are practical gains that free up minutes for school drop-offs, shift changes, and deliveries. A corridor that breathes better also burns less fuel. Operators can retime in hours instead of seasons. Residents notice that the light they used to dread now behaves. The fix looks simple from the driver’s seat, and that is the point.

Ahead of the curve in public health

Outbreaks reward speed. Search data and symptom queries can illuminate patterns before clinics finish their weekly tallies. A landmark study showed how aggregated search behavior tracked influenza-like illness with a lag measured in days, not weeks. That kind of signal helps public health teams plan surge staffing, shift clinic hours, and time messages around handwashing, masks, or vaccines. A forecast can’t administer vaccinations on its own, but it gives the people who lead the response a head start and a set of probabilities they can act on.¹⁴

Open-source intelligence rounds out the picture. The WHO’s Epidemic Intelligence from Open Sources platform scans news, social posts, and official notices to flag early hints of threats in near real time. Analysts validate the signal, connect it to laboratory data, and push alerts through established channels. The workflow turns chatter into decisions that move resources, and it keeps a record of what was seen, when it was seen, and how it was handled.¹⁵ For local health departments, this is a force multiplier. One person in a small office can watch the world, filter it to the county level, and brief leaders before rumors spark panic. The result is calmer communication and faster action when the stakes are high.

Show your work, earn trust

Public institutions win trust when people can see how decisions are made and who is accountable. New York City’s Automated Decision Systems Task Force laid out recommendations to build that capacity inside agencies.¹⁶ The report calls for clear roles, inventories of systems, public documentation, and a way for people affected by an automated result to request an explanation. It also emphasizes records that auditors can follow, version control, and forums where residents can ask questions in plain language. The aim is not a veil of technical detail; it’s a process that surfaces the right information at the right time for the right audience.

Regulators have sketched practical guardrails that teams can apply from day one. The UK Information Commissioner’s guidance frames accountability, transparency, fairness, and rights in terms that product managers and lawyers can both use.¹⁷ It breaks down how to run impact assessments, how to document choices, how to think about statistical accuracy, and how to explain a model’s role in an outcome. That playbook helps agencies choose vendors, structure procurement, and set review checkpoints before launch. The culture shift is simple to describe. Do the work. Write it down. Make the parts that affect people visible and understandable.

AI in public policy delivers wins in speed and consistency, yet big debates follow it everywhere. Three questions come up in hearings, journals, and town halls. Which fairness rule should guide decisions? How much of a model must be explained to the public? When police predict crime, do communities get safer or more surveilled?

Which fairness rule wins when base rates differ?

When groups have different underlying rates for an outcome, some fairness goals clash. Computer scientist Jon Kleinberg and economist Sendhil Mullainathan showed that you cannot satisfy several popular fairness criteria at once when base rates diverge.¹⁸ For instance, you can give every group the same error rates (like equal false positives and false negatives), but then the chance that a “high-risk” label is actually correct will differ by group. If you flip it and make the positive predictions equally accurate across groups, then the error rates end up skewed. 

Their work shows that unless base rates match or predictions are perfect, you have to choose which fairness goal to prioritize. That result forces a choice. Agencies must decide which error to minimize and which equity goal fits the mission. Statistician Alexandra Chouldechova analyzed recidivism risk tools and showed how equalizing one error rate can break calibration across groups, and how calibration can raise disparities in false positives or false negatives.¹⁹

Supporters of the “pick and explain” approach argue that clarity beats vague promises. They want leaders to pick a metric in public, test for harms, and write down the tradeoffs so courts and communities can review them. Critics worry that a single metric invites gaming or hides impacts that show up later in service delivery. The implications are practical. Procurement teams need fairness targets in the contract. Program managers need dashboards that show the chosen metric and its side effects. Oversight bodies need authority to pause systems when unintended harm appears.^18,19

How much explanation do people need, and of what kind?

People want explanations when a model affects benefits, housing, or screening. Legal scholars Sandra Wachter, Brent Mittelstadt, and Luciano Floridi propose counterfactual explanations that tell a person which smallest change would have flipped the outcome, without revealing the full internals of the model.²⁰ The pitch is pragmatic. A borrower sees how to improve. A caseworker sees what to verify. A regulator sees a record that can be audited. 

Legal experts Lilian Edwards and Michael Veale argue that a generic “right to an explanation” in law is too weak and too narrow to fix many algorithmic harms.²¹ They point out gaps in when the right applies, what counts as a meaningful explanation, and how lawyers or engineers would deliver it at scale. The debate shapes day-to-day service design. Agencies need to decide whether to give feature-level reasons, counterfactuals, or full model cards, and which audience each format serves best. A good rule of thumb is to match the explanation to the decision point, keep records, and give a clear path to appeal when something seems off.

Predictive policing: Foresight or feedback loop?

Police leaders see promise in forecasting where crime might cluster. In randomized field trials in Los Angeles and Kent, mathematician George Mohler and collaborators reported that officers using algorithmic hotspots predicted more crime and saw modest reductions in crime volume compared with analyst maps, when measured per patrol time.²² Advocates claim this helps allocate scarce patrols and reduces guesswork during shifts. Critics warn that the data feeding these systems can be tainted by years of biased stops, faulty reports, or undercounted offenses. 

Legal scholars Rashida Richardson, Jason Schultz, and Kate Crawford documented how “dirty data” can lock in past civil rights violations and create feedback loops that send officers back to the same communities regardless of actual harm.²³ The policy stakes are concrete. Cities need transparency on inputs, routine audits for skew, and community oversight where forecasts steer patrols. If a tool is kept, its scope should be narrow, its records public, and its impact on stops and arrests tracked in the open. If those conditions fail, officials should be ready to suspend use and reassess the data pipeline before trust erodes further.

How the VA flagged risk earlier and got people into care

At a Veterans Affairs medical center, a care team starts the week by checking a risk list. A predictive program has scanned records across the system and highlighted a small group of patients who sit in the highest risk tier for suicide at that facility. Clinicians reach out right away. They call veterans, update safety plans, review medications, and schedule follow-ups that fit the person’s routine. The work looks ordinary from the hallway. The key change is the timing—outreach happens before a crisis rather than after one.²⁴

Here is how the program ran at scale. Each month, the system identified the top 0.1% of patients by predicted risk and assigned them to a REACH VET coordinator. The coordinator’s job was concrete. Confirm contact details. Set near-term visits. Make sure a safety plan exists and is in the chart. Check whether therapy, peer support, or telehealth would help. The evaluation used a cohort design with a difference-in-differences strategy. Analysts compared outcomes for high-risk patients after the program began with similar high-risk patients from before the launch, while also tracking trends in patients who were just below the risk cutoff.

What moved the needle was the care activity. Veterans in the program completed more outpatient appointments and had more new safety plans recorded. Hospitalizations for mental health reasons and emergency department visits fell. The most important clinical finding was a reduction in nonfatal suicide attempts over the next six months. Mortality did not change in that early window, which made the message clear for clinical leaders. Keep the high-touch outreach. Keep safety planning high on the checklist. Continue to test ways to reach veterans who face the most lethal means or who engage less with care. The takeaway for any public agency is practical. If a model points to a small, high-risk group, build a playbook that accelerates contact and documents every step so supervisors can see what works and where the team needs help.

Turning online chatter into inspections that matter

Picture a city health department office with two screens glowing. On the left is a feed of new Yelp reviews. On the right is a dashboard that scores each review for signs of foodborne illness. The project team worked with Yelp and university partners to turn public posts into an early warning system.²⁵ Over nine months, software combed through roughly 294,000 New York City restaurant reviews looking for a few simple signals. Keywords like “vomit” and “diarrhea.” Mentions of more than one person getting sick. A gap of at least ten hours between the meal and the symptoms, which helps separate true illness from something like a spicy dish that did not sit well. Each week, an epidemiologist read the flagged posts and decided which ones deserved a closer look.

When a review looked credible, the department sent a private message through Yelp asking the diner for a short phone call. Investigators asked about the date, the party size, what everyone ate, and when symptoms started. They compared the answers to the city’s 311 complaint system and checked for clusters. If multiple credible reviews pointed to the same place around the same time, they opened an investigation. In the pilot, the software flagged 893 reviews. A human screen narrowed that to 468 recent illness posts. Only 3% had also been reported to 311. Interviewing 27 reviewers surfaced three unreported outbreaks that added up to 16 sick customers. Inspectors visited the restaurants and found violations like poor cold storage, cross-contamination, and evidence of pests.

The message is straightforward. Online reviews will never replace lab tests or routine inspections, but they can widen the net. A post written on a Sunday night can prompt a call on Monday and an inspection later that week. Residents see faster action. Inspectors extract better leads that focus scarce time. City leaders get a program they can explain in a single slide. Pull what the public already shares. Score it with simple rules. Have a human read the edge cases. Keep a record of what the team did and what changed. Secure trust between vendors and customers and promote health at the level of the community.

Algorithms for Simpler Decision-Making (1/2)

This piece explores how our minds extend themselves into digital tools when algorithms support memory, analysis, or suggestions. Readers discover how decision support systems can serve as cognitive prosthetics, helping us tackle complex problems like financial planning or medical triage more effectively. You’ll also learn how policymakers use these tools responsibly, balancing speed with user agency.

The "Mystery" of Intuitive Decision Making

This article celebrates the power of gut feeling informed by experience. It reveals how seasoned professionals, from expert chess players to financial advisors, can make swift, reliable judgments with minimal information. The piece explores psychological research that compares intuition and analysis, showing when a well-honed instinct can outperform slow, data-driven deliberation. You’ll come away with an appreciation for blending human judgment with algorithmic insight for smarter, faster choices.

1. National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0) (NIST AI 100-1). https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf
2. Treasury Board of Canada Secretariat. (2024). Directive on Automated Decision-Making. https://www.tbs-sct.canada.ca/pol/doc-eng.aspx?id=32592
3. Kleinberg, J., Lakkaraju, H., Leskovec, J., Ludwig, J., & Mullainathan, S. (2018). Human decisions and machine predictions. The Quarterly Journal of Economics, 133(1), 237–293. https://doi.org/10.1093/qje/qjx032
4. O’Neil, C. (2016). Weapons of math destruction: How big data increases inequality and threatens democracy. New York, NY: Crown.
5. McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (1955). A proposal for the Dartmouth summer research project on artificial intelligence. https://doi.org/10.1609/aimag.v27i4.1904
6. Samuel, A. L. (1959). Some studies in machine learning using the game of checkers. IBM Journal of Research and Development, 3(3), 210–229. https://people.csail.mit.edu/brooks/idocs/Samuel.pdf
7. Executive Office of the President, National Science and Technology Council. (2016). Preparing for the future of artificial intelligence [PDF]. https://obamawhitehouse.archives.gov/sites/default/files/whitehouse_files/microsites/ostp/NSTC/preparing_for_the_future_of_ai.pdf
8. Eubanks, V. (2018). Automating inequality: How high-tech tools profile, police, and punish the poor. New York, NY: St. Martin’s Press. Colorado Mountain College
9. Buolamwini, J., & Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in commercial gender classification. In Proceedings of the Conference on Fairness, Accountability, and Transparency (FAT*) (pp. 77–91). https://www.classes.cs.uchicago.edu/archive/2020/winter/20370-1/readings/gendershadesAIbias.pdf Department of Computer Science
10. Organisation for Economic Co-operation and Development. (2019). Recommendation of the Council on Artificial Intelligence (OECD Legal Instruments, OECD/LEGAL/0449). https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449
11. European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act). Official Journal of the European Union, L, 12 July 2024. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng EUR-
12. Radin, S., Chajka-Cadin, L., Futcher, E., Badgley, J., & Mittleman, J. (2018). Federal Highway Administration Research and Technology Evaluation Final Report: Adaptive Signal Control (FHWA-HRT-17-007). U.S. Department of Transportation, Federal Highway Administration.
13. Skabardonis, A. (2013). Advanced traffic signal control algorithms (CA13-2157-A). California PATH, University of California, Berkeley. California Department of Transportation.
14. Ginsberg, J., Mohebbi, M. H., Patel, R. S., Brammer, L., Smolinski, M. S., & Brilliant, L. (2009). Detecting influenza epidemics using search engine query data. Nature, 457(7232), 1012–1014. https://doi.org/10.1038/nature07634
15. World Health Organization. (2025). Epidemic Intelligence from Open Sources (EIOS) initiative. Geneva, Switzerland: WHO.
16. New York City Automated Decision Systems Task Force. (2019). New York City Automated Decision Systems Task Force report. New York, NY: NYC Mayor’s Office of Operations.
17. Information Commissioner’s Office. (2023). Guidance on AI and data protection. Wilmslow, UK: ICO.
18. Kleinberg, J., Mullainathan, S., & Raghavan, M. (2016). Inherent trade-offs in the fair determination of risk scores. arXiv preprint. https://doi.org/10.48550/arXiv.1609.05807
19. Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), 153–163. https://doi.org/10.1089/big.2016.0047
20. Wachter, S., Mittelstadt, B., & Floridi, L. (2018). Counterfactual explanations without opening the black box: Automated decisions and the GDPR. Harvard Journal of Law & Technology, 31(2), 841–887. https://doi.org/10.48550/arXiv.1711.00399
21. Edwards, L., & Veale, M. (2017). Slave to the algorithm? Why a “right to an explanation” is probably not the remedy you are looking for. Duke Law & Technology Review, 16(1), 18–84.
22. Mohler, G. O., Short, M. B., Malinowski, S., Johnson, M., Tita, G. E., Bertozzi, A. L., & Brantingham, P. J. (2015). Randomized controlled field trials of predictive policing. Journal of the American Statistical Association, 110(512), 1399–1411. https://doi.org/10.1080/01621459.2015.1077710
23. Richardson, R., Schultz, J. M., & Crawford, K. (2019). Dirty data, bad predictions: How civil rights violations impact police data, predictive policing systems, and justice. NYU Law Review, 94, 192–233.
24. McCarthy, J. F., Cooper, S. A., Dent, K. R., Eagan, A. E., Matarazzo, B. B., Hannemann, C. M., Reger, M. A., Landes, S. J., Trafton, J. A., Schoenbaum, M., & Katz, I. R. (2021). Evaluation of the Recovery Engagement and Coordination for Health-Veterans Enhanced Treatment suicide risk modeling clinical program in the Veterans Health Administration. JAMA Network Open, 4(10), e2129900. https://doi.org/10.1001/jamanetworkopen.2021.29900
25. Harrison, C., Jorder, M., Stern, H., Stavinsky, F., Reddy, V., Hanson, H., Waechter, H., Lowe, L., Gravano, L., & Balter, S. (2014). Using online reviews by restaurant patrons to identify unreported cases of foodborne illness: New York City, 2012–2013. Morbidity and Mortality Weekly Report, 63(20), 441–445.https://www.cdc.gov/mmwr/preview/mmwrhtml/mm6320a1.htm